> ## Documentation Index
> Fetch the complete documentation index at: https://docs.connie.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Work in progress

> See what an agent is working on and for how long, and a conversation's tasks, in Chat details or on the In progress page; follow what it is doing while it works, stop work, approve actions in a room, and get notified when work needs you.

> [!NOTE]
>
> Native Task release conditions are pending rollout. Until activated, requests to save these conditions are refused rather than silently saved as description text.

For a Task that must wait for payment or written approval, ask the agent to bind release conditions to the exact saved Record Fields. For example: “Require Paid amount to be at least Quoted amount and Written licence approved to be Yes before this release Task starts or completes.” The agent needs access to those sources; it cannot substitute a claim in its reply for a saved value. Current unmet or unavailable conditions block status progression and delegated work. Read the Task's **Release conditions** to see which checks still need attention.

> [!NOTE]
>
> To track work that people and agents share, with an owner, a status and a review, use the [Tasks](https://docs.connie.ai/tasks/overview.md) page. This page is about the work an agent is doing right now, which is never a task by itself. In a conversation, Chat details lists both: its work under **In progress**, its tasks under **Tasks**.

> [!NOTE]
>
> **Before opening Connie Settings:** if pending Chief of Staff setup sends you back to chat, finish the setup cards there, then reopen Settings. Direct Settings links follow the same check. New workspaces use free credits, with no card or trial. **Billing** (`?tab=plans`) remains available during setup. See [Settings access during setup](https://docs.connie.ai/workspace/settings.md#if-settings-takes-you-back-to-your-chief-of-staff).

When you ask an agent for something, it starts working on it. That work is **in progress** until it's done, stopped or failed, including while it waits for an approval, a connection or your answer. Work an agent [runs on its own](https://docs.connie.ai/agents/agents-and-permissions.md#work-it-runs-on-its-own) is in progress the same way, without anyone asking: that's **background** work. Work in progress isn't a task: a task is something you track in [Tasks](https://docs.connie.ai/tasks/overview.md), and an agent can work on one.

One agent can work on independent tasks at the same time, including in the same room or across rooms. A follow-up steers your existing work; ask for a separate task to start another assignment. If your message could refer to several tasks, the agent asks which one. Messages you send within a few seconds of each other count as one request: they go to the same work and you're never asked about them (see [Worked for and tool details](#worked-for-and-tool-details)). Each task keeps its own answer, decisions and Stop control. A saved browser account can still be controlled by only one task at a time.

If a temporary request limit or busy workspace delays the start, your work stays queued and retries automatically when capacity is available. You do not need to send it again. Queued work stays quiet until it starts, and you can still stop it. Permissions and available credits are checked again before it runs. The notice under your request explains a saved waiting reason: **Queued because a request limit was reached. \[agent] will retry automatically.**, **Queued while this workspace's other requests finish. \[agent] will retry automatically.**, or **Queued because Connie couldn't start this request yet. \[agent] will retry automatically.** Older waits with no saved reason say **Queued. \[agent] will retry automatically.** A queued request does not mean the agent itself is busy.

Agents reply in a friendly, upbeat tone with short paragraphs, usually one to three sentences. Clear requests go straight to work without a routine plan card. If decisions block progress, the agent asks them together on one short question card. Ask for more detail whenever you need it; requested reports and documents remain complete. Approvals, paid-run quotes and access restrictions still apply.

## Checking results

Agents check saved results before handing them over. For a List or data change, this means reading back the saved Fields, checking counts, duplicates and required missing values, and comparing a sample against your criteria and exclusions. A successful run or a “Great” label alone does not prove a Record qualifies. The reply should match the saved results and identify any shortfall or check that could not be completed.

When checking a finite set is inexpensive, the agent checks every item before giving a verdict for the group. A partial check says how many items were checked and what remains unverified. Rejection reasons must follow your actual criteria; unavailable evidence means unconfirmed, rather than a proven reason to exclude something.

## In progress and Routines

When one agent asks another for work, the receiving agent can send a result or a blocker straight back. Its message wakes the original sender in the conversation where the request began; you do not need to relay reports or set up a polling routine. If the sender is still working on the request the reply answers, the reply waits for that work instead of starting a second answer beside it: when the sender is waiting for the answer, it reads the reply there and you get one closing message; otherwise the reply wakes it once that work ends. If someone stopped that request, the late reply starts nothing. Replies keep the original requester, permissions and conversation access. Stopping the work prevents late replies from restarting it, and paused agents receive no new agent handoffs. New delegation remains limited to two hops; a correlated exchange permits up to 16 consecutive replies to prevent loops.

This works in private agent conversations and between agents present in a shared room, including saved requests that used the older messaging tool. Existing isolated helper conversations outside a shared room keep their earlier behavior and do not gain this direct return notification.

Event menus in agent configuration use the available space above or below their field. On short windows or at high zoom, scroll inside the menu to reach the remaining events.

Click the centered agent or room name at the top of the conversation to open **Chat details**. The panel contains **In progress**, **Tasks** (in workspaces with the [Tasks](https://docs.connie.ai/tasks/overview.md) page), **Routines**, **Artifacts** and **Context**. Close it with **Close details**, **Esc**, or by clicking the centered name again.

**In progress** shows live work. While the work is going on, its active work time ticks. When it has a status such as **Needs connection**, only that status appears. Its icon matches the responsible agent’s color. **Needs connection**, **Needs approval** and **Waiting for reply** appear as chips in that agent’s color: click one to scroll to its card in the conversation. Other statuses stay gray. It disappears when there is nothing in it. While work is being read, you see **Loading…**; a failed read shows an error instead of claiming the list is empty. Finished work leaves the list, and a failed run stays until its failure has been delivered.

**Tasks** lists the conversation's tasks from [Tasks](https://docs.connie.ai/tasks/overview.md): the tasks that belong to it, the workspace's tasks an agent worked on here and, in your DM with an agent, the tasks delegated to that agent that you can see. Each row has the task's status, its title and its identifier, such as **LUM-4**; what waits on a person (**In review**, **Triage**) comes first. Every open task is listed, however many there are. Closed tasks (**Done**, **Cancelled**, **Duplicate**) are listed too while there are four or fewer; past that they wait behind **Show N done**. While an agent works on one of these tasks in this conversation, the task moves to **In progress**, shown with its identifier and its live status, and comes back to **Tasks** with its new status when that work ends. Click a task, or the identifier on an **In progress** row, to open it in place of the details; **Back to chat details** or **Esc** returns. If you can post in the conversation, **+** beside **Tasks** (**Add task**) creates a task associated with this conversation; like every Task, every non-guest member of the workspace can see it. You can also ask an agent to keep track of something: it creates the task, and it tracks its own work as a task when the work goes beyond its reply, for example something it will come back to. In workspaces without the Tasks page there is no **Tasks** section.

**Routines** keeps saved schedules, watches, reviews and event triggers between runs. Each row has a short generated title, a simple description of when it runs, and an enable/disable switch in its responsible agent's color. Event-triggered routines have a lightning icon; scheduled routines keep repeat arrows. The same icon appears when you open the routine's details, including when it is paused. The list icons have no hover tooltip. Timed routines keep their time zone, for example **Every day at 8:00 AM (Asia/Makassar)**. Hover over a truncated title or focus it with the keyboard to read the full title. Shared rooms also identify the responsible agent; agent DMs omit that redundant tooltip. A running occurrence appears separately in **In progress** while its definition remains in **Routines**. To run a scheduled routine once, ask the agent in chat; its automatic schedule stays unchanged.

You can ask your agent to create a routine for a schedule or supported event, such as “When a new email arrives, draft a reply and DM the draft to me on Slack.” The agent can discover the available triggers and required settings, then save the routine in this conversation. Choose the source account when needed. New routines save off, except a routine an agent creates for itself to look after a goal it owns, which starts enabled (see [Work stays open until it is done](#work-stays-open-until-it-is-done)). If you ask for work to run on its own and then approve its plan, that approval also asks the agent to enable the routine. It verifies the saved active state before confirming. Tool approvals still apply. The same source access, approvals and audience checks apply as in Chat details.

Ask the agent to read back a routine’s full saved instructions to verify them. Saved instructions are returned up to the full 4,000-character limit. A read-back does not run, enable or change the routine.

Trigger settings come from the shared catalog. Calendly bookings can exclude rescheduled meetings. HubSpot updates can watch one changed property, and updates to another property do not start the routine. Account choices stay in the authorized account picker.

New assigned routines are **off** when saved, including routines created in chat. They have no automatic next run until enabled. Existing routines keep their saved state.

To add a routine, click **+** beside **Routines** (**Add routine**): it starts "Create a routine: " in the message box, and you tell the agent what to do and when. Open a saved routine to change its trigger. **Schedule** is the first trigger option: choose an interval in minutes or hours, a daily or weekly time, or a custom weekday schedule. Scheduled runs must be at least 5 minutes apart, including custom schedules. Shorter intervals show a message beside the field and keep Save disabled. Times use the selected time zone. Search the event list by name; events needing an unconnected app are hidden until you select **Show all events**. No event is preselected. Choose an agent you own in this conversation, describe what it should do, then choose a schedule or event and any account or conditions it needs. Its response appears in the agent DM or room where you created the routine. Only an agent's owner can add or change its event routines. Other readers can see the routine's title and trigger description. Its source accounts must remain available to its owner and allowed in the agent's permissions. If the room's access changes, turn the routine off and on again to confirm its current audience before more events can start work.

Each matching event runs the routine's saved instructions against that event. Include the details it needs in those instructions; an event run does not continue unrelated requests or pending approvals from the conversation. Each event has its own result, even when earlier events produced similar replies.

Agents can also react to other agents without a scheduled check. Ask for **Agent Task Completed**, such as “When Sarah finishes a task, review her result,” or **Agent Message Received**, such as “When Sarah sends you a report, check its claims.” Choose a specific source agent or any other agent in the workspace. The reacting agent’s owner can ask the agent to create this routine; it stays off until explicitly enabled. Completion events include the saved result. A received-message routine guides the task that already receives that message, so it does not start a duplicate run. The result appears in the chat where the routine was created.

An event does not grant access to another agent’s private chats. Everyone who can read the destination must be allowed to read the source result. Event reactions also stop before returning to an agent already in the same reaction chain. Retries deliver the same occurrence once; enabling a routine does not replay earlier work.

Scheduled occurrences and matching events return an update each time; hourly schedules are not capped at six updates. Watches and reviews share a limit of six new posts per agent in a chat over 24 hours; their checks may continue after that limit. Quiet hours and access checks still apply.

During setup, expand **Mailbox and Slack access** to inspect available accounts. Reading a teammate’s mailbox does not allow sending: its owner or an admin must enable **Teammates can send from it** in Plugins, and the agent must still have account access. Inspect Slack channels separately for each connection; a channel visible to one account may be unavailable to another, and Slack posting permissions still apply.

Right-click a routine card to **Pause** it or **Delete** it. Paused routines offer **Resume**. With the keyboard, focus the title and press **Shift+F10** or the context-menu key. You can also ask in chat, such as “delete my daily update routine”. The agent checks the saved routines and asks which one if the request is ambiguous. The same permissions apply in chat and on the card. Deleting removes future runs and stops linked work while keeping history and completed actions.

### Run a routine once now

Ask the agent to "run it now" to start one run of a scheduled, watch or review routine straight away. It runs exactly as a scheduled run does: on its own, with the saved instructions, on the same accounts and permissions, and it posts where the routine reports. It does not change the schedule or the next run, it also works for a routine that is still off, and it does not start while an earlier run of the same routine is in progress. Agents use the same run to check a routine they just created before you rely on it. When the run finishes, the agent tells you what it posted, or that it posted nothing because it found nothing new. If that run needs something only a person can give, such as an approval or a connected account, the agent tells you exactly what. A run on its own does not show connect cards or questions; it reports what it needs instead. Approvals given with **Allow in this session** belong to one conversation, so they apply to a routine's runs only when it reports into that same conversation; **Always allow** applies everywhere. An event routine runs only when its event arrives. A run now uses credits like any scheduled run.

Only a person can ask for a run now. Work that runs on its own, such as a routine's scheduled run, a follow-up, a later look at unfinished work, work a routine handed on, or an agent picking up what happened (see [Agents pick up what happens](#agents-pick-up-what-happens)), never runs a routine, turns one back on, deletes one or creates a new one: if a routine failed, the agent tells you what happened, and you can ask it to run the routine again. A routine's own run can still pause that routine when its job is done, and it can update its own instructions or notes, or move its own schedule later or less often, but never make a routine run sooner or more often, change its quiet hours or priority, or change another routine. It also never turns a paused agent back on, pauses another agent, or creates or deletes agents or their triggers. Work it hands to another agent, including a Task it assigns or a message from a scheduled Connie chat, runs on its own too, with the same limits, and a paused agent is not handed it.

Turning a routine off keeps its definition and stops its linked work. If saving fails, the switch keeps its saved value and shows an error so you can try again. A paused agent does not run its routines until it resumes. The **Artifacts** and **Context** sections stay available even when there is no work in progress.

If you ask to pause a routine until a specific time, its card in **Routines** and its opened details show the saved resume date, time and time zone, for example **Paused by Yann until Mon 9 Nov, 09:00 (Europe/London)**. Its switch stays off during the pause. A pause without a resume time has no automatic resume; choose **Resume** when you want it to run again.

Created, updated and triggered routines use a compact centered clock-and-title line in the chat. Every run of a routine that posts in its room or agent DM, whether it ran on its schedule, on its event or because you asked to run it now, starts with a **Triggered routine** line naming the routine, above what it posted. A run that finds nothing new and posts nothing leaves no line; a run that could not finish keeps its line, with "This run couldn’t finish, so nothing was posted." under it.  Triggered occurrences have no separate date line; normal conversation date separators remain.

A scheduled occurrence returns an update to its room or agent DM when it has something to say: what it did, what changed, what needs you, or why it could not finish. When it did its work and nothing changed and nothing needs anyone's attention, it posts nothing, so you hear from a routine only when there is something to read. If you want a report every time, say so in the routine's instructions (for example “send the summary every morning even when nothing changed”) and it writes one each run. A failed run notifies the person who set it up. Repeated results and the limit on unsolicited watch updates do not hide scheduled replies. Quiet hours hold scheduled updates until they end. Watches, reviews and event routines stay quiet the same way when there is nothing new to report. A routine's update contains its report only, not the notes the agent made to itself while it worked.

Each **In progress** row shows the work (or the task it's for) and either how long the agent has actively worked on it or, when it waits, only what it's waiting for:

| Row | Meaning |
| - | - |
| `45s` | Working now; the time ticks while it works. |
| `Needs approval` | Waiting for someone to approve an action. |
| `Needs connection` | Waiting for an account to be connected or reconnected. |
| `Waiting for reply` | The agent asked a question and waits for your answer. |
| `Waiting until 9:00 AM` | Waiting for a time. It can also wait for an event or for another agent (`Waiting for [agent]`), or just say `Waiting`. Work that is not finished yet shows what it waits for here, with a **Left:** line under its title (see [Work stays open until it is done](#work-stays-open-until-it-is-done)). |
| `Stopping` | Being stopped. |
| `Queued` | Waiting to start, for example behind another person's request to the same agent in the same room. |
| `Checking outcome` | An action may have run but its result wasn't confirmed; it is being checked, never simply repeated. |

The time counts only active work, never the waiting, and it's the same after a reload and on every device. Finished or stopped work leaves the list; its result stays in the conversation. Work whose run ended before it was finished stays in the list. Failed work stays until the failure has been reported in the conversation. The agent also keeps a bounded record of its earlier tool outcomes in this conversation, including failed attempts and work waiting for approval. A failed or unapproved step is never treated as completed work.

If an agent couldn't start on a request because of a problem on our side, the conversation says "\[agent] could not start on this because of a problem on our side. Try again in a moment." with a reference to give support; if it was refused for a reason you can act on, such as running out of credits, it says that instead. A request that stops later for the same kind of reason is never shown as failed.  When the agent could save a reply, it says in its own words what it finished, what stopped it and what to do next. When the run ended without saving one, the conversation says "This stopped on our side before it could finish. Ask again to start it over." after the agent's name, sometimes followed by a reference to give support.

When the AI provider is overloaded or stops responding during a step, the agent tries that step again by itself after 15 seconds and again after a minute, and a step cut off by a restart on our side is picked up again; no step that already ran is repeated, and you see the usual activity line meanwhile.  If an agent's work is cut off and its run ends without saving a result, it is not started again: the conversation says "This stopped on our side before it could finish." after the agent's name. When only reads had run, it adds "Ask again to start it over." When a step that changes something, such as renaming a List or sending an email, had already started, it names that step, says it may have taken effect and adds "It was not run again, so check its result first."; the agent never runs that step again by itself. A reply the agent had already shown for that request is removed then, so the request leaves one message. Nothing in a room or agent DM is shown as failed, and there is no **Try again** button there.

Click a row to scroll to its card in the conversation, or to the request it is working on. You can also focus its title and press Enter or Space. Hover a row, or move to it with the keyboard, to see **Stop** (always shown on a touch screen). The title’s tooltip names the responsible agent.

Work you asked for in Slack shows the same way in the agent's profile, with a Slack mark that opens the thread; its answer stays in Slack. See [Agents in Slack](https://docs.connie.ai/agents/agents-in-slack.md).

While an agent uses a browser, its interactive card appears above the message box. When a browser needs your help, the **Computer** card shows **Action needed** and the instructions for that step. **Take over** opens the same browser fullscreen, requesting control first if the agent is still using it. The bottom control shows **You’re in control** and **I’m done**. **Exit full screen** returns to the conversation without handing control back or reconnecting the browser. Escape also exits while keyboard focus is outside the website; when you are interacting with the website, use **Exit full screen**. Choose **I’m done** to return control: waiting work continues, and when the agent had asked you for a step and stopped, **I’m done** posts “I've finished in the browser. Please continue.” in the chat so it carries on. **Skip** appears on the inline card only when an agent has browser work waiting for your answer; it saves the account and tabs and cancels that work. Taking over a running browser yourself does not show **Skip**. In the side panel, use **Save and close** to close the browser or cancel waiting browser work. Once the answer is published, its browser card stays in the conversation; you do not need to open **Artifacts** for its controls. The agent can scroll long pages and nested scrolling panels to read beyond the first screen. When it needs a website sign-in or verification, it pauses with a browser card in the conversation. Only the owner of the browser account can interact with it. Private browser work requires a conversation whose readers are limited to that owner; adding room readers does not share the browser. Where a saved handoff is visible to other readers, it says whom it waits for without showing the private session. Complete the step inside the browser and choose **I’m done**. Extra controls, including **Import cookies**, **Expand browser** and **Save and close**, are under **Browser actions**. Passwords and verification codes stay out of the conversation. A reader who cannot post cannot finish the step. If the embedded viewer fails, choose **Reconnect viewer**. While you have control, **Open browser in new tab** also opens the same live session separately; return to the card and choose **I’m done** when finished. See [Browser accounts](https://docs.connie.ai/workspace/settings.md#browser-accounts).

When an agent creates or edits a routine, a centered **Created routine** or **Updated routine** line appears in the conversation with a clock icon and the routine’s saved title. Click the line to open the routine in Chat details. These lines also appear when you reopen the chat. Failed saves leave no success line.

If you name a routine when you ask for it (for example “a routine called Weekly pipeline check”), it is saved with exactly that name and keeps it. Otherwise the agent gives it a short name of a few words describing what it does, before it is saved, so the name the agent tells you is the same one shown in **Routines**, in **Created routine** and in later **Triggered routine** lines.

When you ask an agent which routines it has, it also lists the routines it runs in other conversations you can read, by that conversation's name and with a link to it.

## Files in Context

**Context** lists the files people have shared in this conversation, including files from older messages. Each row uses an icon for its file type, such as a spreadsheet, image, PDF, audio or video. Click a file to open it beside the conversation in the file viewer. Point at a file for a short description derived from its readable text or file type. The same description is available to screen readers. Use **Load more files** for older files. Files the agent creates appear under **Artifacts**. File access is checked again when you open it.

An upload joins Context when you send its message. Context stays available when an agent condenses a long conversation: it keeps a small file catalog and reopens the relevant files when needed, rather than repeating every document in each reply. Deleting a message or using **Edit and resend** removes files from Context when no remaining message shares them. Files stay within this conversation’s access; uploading one does not make it available in other rooms.

Use **Call** to start or join an [audio call](https://docs.connie.ai/agents/rooms-and-people.md#audio-calls) with the room’s people and agents. **Slack** is shown disabled while it is unavailable. Ask an agent in the conversation for work, and attach files from the message box. If you can post in the conversation, **Add artifact** (**+** beside **Artifacts**) starts "Create an artifact: " in the message box, ready for you to finish and send; it is not offered in a read-only conversation.

Spoken work requests use the same work in progress, credit and plan requirements, permissions, connections and approvals as typed requests. If a plan is required (for example because the free credits are used up), the agent explains that no work started and asks you to choose a plan in the app. The agent can keep talking while the work runs and read its concise published result back into the call. Spoken requests keep their speaker and exact words on the work without posting a duplicate human message. Call transcripts appear in **Chat details → Context**, and relevant clarifications remain available when work resumes after approval. You can ask for a saved Records List of the actual collected items rather than a count summary or Page. Asking an agent to leave the audio call does not cancel its work; stop it with **Stop** in **In progress**.

## Open an artifact

The card's **Artifacts** lists what the conversation made or uses: pages, files, apps, Records results, Lists and Workflows. Click one to open it. One that was deleted since, or that you can't open, says **Unavailable** and doesn't open. Each one opens beside the conversation, as it does beside a chat, and you stay in the room: a page, a file, an app, a saved Records result, and a Workflow, a List or an Object too, with the same controls it has on its own page. Clicking another artifact replaces the open one, **Close** at the top right of its header closes it, and **Full screen** expands the same panel in place. **Exit full screen** or **Esc** returns to the side panel. You only see what you already have access to. An artifact card's **Open** in the conversation works the same way. A Lead Finder card, a search an agent prepared for you (such as "B2B SaaS decision-makers"), opens the [Lead Finder](https://docs.connie.ai/records/lead-finder.md) over the conversation with the agent's filters and previews the matches, as it does in a chat; nothing is imported until you import, and the List you import into then opens beside the conversation.

In rooms and agent DMs, **Open** on a workflow, analytics view or app card also keeps the conversation beside the artifact. This works for saved replies as well as new results; you can keep reading and writing in the conversation while the artifact is open.

A link to an artifact in a conversation (your copy after **Duplicate**, for example) opens that conversation with the artifact beside it. You only ever see artifacts the conversation lists for you: a link to one you can't open there opens nothing.

## In progress

While an agent is working on something you can see, its face in the sidebar's **Agents** area slowly looks around; a screen reader hears "working" in its name. The face stays still if your device is set to reduce motion, and while the agent only waits (for an approval, a reply or a time). The sidebar has no **In progress** entry: the page with all of it in one list is at `https://connie.ai/<workspace-id>/rooms/tasks`. It says: "What your agents are working on, in every room and conversation you can see. Finished work leaves this list; its result is in its room."

Each row in this separate overview shows the work, the agent, where it reports and how long it has worked (for example `2m · Needs approval`). Click the room's name to open it. Work in a room you're not in shows "A conversation you are not in", never its name; work you asked for from Slack or from Connie in your AI apps shows **Slack** or **Connie**. Background work, from a responsibility or a schedule, is marked **Background**. **Stop** appears where you may stop the work. With nothing in progress, the page says "Nothing in progress."

## Stop work

* **Stop** on a row stops that work at once. The button shows **Stopping…**, the row says **Stopping** until the work has stopped, and then it leaves the list. If the work was for a Task that is **In progress**, the Task goes back to **Backlog** and waits there until someone moves it to **To do**; stopping work never closes a task.
* You can also tell the agent in the conversation, while it works, to stop. If the conversation has several agents and it isn't clear whose work you mean, you're asked which agent you mean, and nothing is stopped until you say. One of the agents asks in its own words; if that agent is busy itself, the conversation asks at once instead, naming what each one is working on, for example "Which agent is this for? Chief is working on "Search for Acme contacts"; Scout isn't working on anything here. Mention the one you mean." Mention the one you mean, for example **@Scout**, and that agent stops the work your earlier message meant. If that work had already finished by the time you answer, the agent says so and that nothing was stopped. The same goes for any message you send while the agent works that reaches it only after that work ended: it says the work had already finished (or had been stopped, or had ended without finishing) and that nothing was stopped.

In your DM with an agent, **Edit and resend** on one of your messages also stops the work that message and the later ones started (see [Rooms and people](https://docs.connie.ai/agents/rooms-and-people.md)).

Only the person who asked, the agent's owner and the room's non-guest participants see **Stop** on a row. Stopping can't undo anything already done, such as an email already sent, and no new step starts after it. If the work was waiting for you in a browser, that browser is saved and closed with it: your login and tabs are kept and the card offers **Reopen browser**. A step already under way when you stop, such as a web search, is stopped if it can still be, and is charged only for what it did before the stop; one that finishes anyway is charged as usual. Work the agent had handed to other agents for it stops too, with anything they had handed on in turn, also below work of theirs that shows **Checking outcome**, which is still checked as usual. Stopping work skips what remains of it without an error or a "Response was interrupted" card, including after a reload. Anything it made, such as a Page or a List, keeps its card. A run with no visible steps leaves no empty bubble or summary control. When you ask the agent to stop in the conversation, its reply says what it had already done before it stopped, including what another agent it had handed part of the work to had done. What you wrote stays in the conversation: in the other agent's Activity, its stopped work reads "The work that handed this on was stopped." If steps were under way when you asked, it waits a few seconds for them first: it names those that ran, says which it stopped before they finished, and says when it can't tell yet whether one finished. It names that agent, and the steps it took, only if everyone who reads the conversation may know of it: the agent takes part in the conversation, is shared with the whole workspace, or, in a conversation only you read, is one you can see. Otherwise the reply says only that another agent had done some of the work, or was in the middle of a step.

## Worked for and tool details

Incoming messages in agent DMs and rooms never show a **Worked for** duration, including historical replies. Replies have no **Steps**, **Stopped** or duration summary control and reserve no space for one. Tool calls are never listed in the conversation: while an agent works, its activity line says what it is doing (see [Who answers](https://docs.connie.ai/agents/rooms-and-people.md#who-answers)), and its finished reply says what it did.

While an answer waits for an approval, a connection or a reply, its card stays in view under the agent's message. Once the card is approved, it leaves and the action runs with the agent's other activity; the agent's reply says what it did.

An action someone rejected does not run. The agent's following reply explains the outcome, including after reopening the conversation, without a separate execution row or an empty bubble for a work-time line. If someone posted a message while the action waited, the earlier messages stay above it and the continued reply stays below it (for stopped work, see [Stop work](#stop-work)). A room's notices, such as "\[agent] was removed from this room, so it stopped working on this.", remain separate from answers.

For a new work request, the agent can first post a short sentence saying what it will prepare or check. When you ask for a Workflow dry run or a test that must not send, this opening describes checking whether a safe simulation is available; it does not promise that a simulation will run. Corrections and requests to use a different source or tool get this same opening, including when phrased as a question. Quick factual questions, exact-output instructions and Stop do not need an opening. An answer appears in the conversation once it is finished. While it is being written, the activity line says so (for example **Writing** or **Putting it into words**). Each request gets its own answer and tool activity, even when several agents work at the same time or two people ask the same agent one after the other.

If you send another message while an agent is working on your request, such as a correction or something to add, it joins that work instead of waiting for it to finish. The agent reads it before its next step, changes course if it needs to, and its one answer covers both messages; your later message is not listed as a separate task. The same goes for a request you send as several messages in a row, such as a brief pasted line by line: messages sent within about 15 seconds of your previous one to the same agent join the work that message went to, in one run with one answer, and no **Which task is this for?** question is asked about them. To start separate work beside it, say so ("separately", "in parallel") or send it a little later. If it arrives while the answer is being written, the agent writes the answer again with it. A message asking the agent to stop still stops the work (see [Stop work](#stop-work)). In your DM with an agent, **Edit and resend** on a message that joined its work stops that work too, as it does the work any edited message started. Another person's message to the same agent is a request of its own. When a request is waiting behind another run, it shows **Queued. \[agent] starts on this when it finishes what it is working on now.** under it until then. A wait caused by request limits or workspace capacity has its own notice instead (see above).

Approvals and questions stay visible as cards. An answer whose step made a Page, a List or another artifact shows it as a card under the answer.

In agent DMs and rooms, live work appears in the activity line above the message box, with a separate line for each active request. Tool calls never appear as execution rows in the conversation, including after a reload. Approval and question cards remain visible when a decision is needed, and artifact result cards stay in the conversation. Open **Chat details → In progress** for task status and controls; select a task to return to its request or waiting card.

## Work stays open until it is done

Work is finished when what you asked for exists, not when the agent stops typing. Before a reply is posted, a separate review compares your request with what the agent actually did (see [Chat](https://docs.connie.ai/chat/overview.md)). The same review decides what the work becomes:

* **Done.** Everything you asked for was done. The work leaves **In progress**. Runs the agent started and saw finish during its reply count as done, including a run you approved on a quote card. If the agent tells you plainly that something can't be done from where it is (it has no access, the item isn't in this conversation, there is no tool for it) and what would unblock it, that is also a finished reply: the agent isn't sent off to look for a way round. An offer of a next step at the end of a reply doesn't keep the work open either.
* **Not finished, and the agent can go on.** The agent is sent back to work, up to twice in the same reply, and not again for a reply it keeps unchanged. If it still isn't finished, the reply says what is left, and the work stays under **In progress** with a **Left:** line under its title and the time of its next look, for example `Waiting until 3:40 PM`.
* **Waiting on something.** The work stays under **In progress** and says what it waits for, for example `Waiting for your answer on which Calendly link to use`: an answer or decision of yours, an approval, an account to connect, a date, or something outside Connie.

A decision you made is not something the work waits for. If you reject an approval card, or decline to connect an account, the agent tells you what was not done and the work leaves **In progress**: it is never kept waiting for a card you already decided. Work that was waiting for that approval or connection ends too, and a later reply never asks you to approve a card you rejected. If you change your mind, ask again and the agent asks for approval again. Likewise, when you asked for something another run carries out (for example "run this routine once now") and your request waits on that run's approval card, your request leaves **In progress** as finished once you approve the card and that run has finished; it doesn't stay waiting for a later check.

The work's title names your whole request. A brief with several parts is titled as one piece of work, such as **Website brief: 8 changes**, never as two of its parts.

**Open work is picked back up without a reminder.**

* When you send another message in the same conversation, the agent answers it and then continues the open work in the same reply, or tells you plainly that it is parked, what is left and what will resume it. Your new message is still its own work with its own row.
* If nothing arrives, the agent looks at the open work again on its own: at the time it named, or otherwise after 30 minutes, then 4 hours, then a day. Something new brings that look forward, such as a run finishing that the agent had not yet seen finish; a run whose result the agent already reported does not. It works in the background and writes to you in the same conversation only when there is something to say: the work is finished, part of it got done, or there is something only you can settle. A look that finds nothing new posts nothing. If a look needs a decision from you (an approval, an answer or a connection), its card appears in the conversation where the look stopped, and **Needs approval** in **Chat details** scrolls to it.
* Before a look runs, the platform checks what changed since the work was last looked at: new messages in the conversation, a run the work started finishing, work it handed to another agent ending. When the agent is waiting on a run that is still going or an agent that hasn't answered, or when its last look found nothing and nothing changed since, the look is made without running the agent: no credits, nothing posted. It still counts as one of the work's looks, and the next one follows the same schedule. When something did change, the agent looks as usual and is told what changed. A look at work waiting for a date or a time, and the first look at work the agent has not finished, always runs the agent.
* After three looks that did not finish it, the work ends and leaves **In progress**. Nothing is posted if the last look had nothing new. The agent keeps what was left: ask about it, or ask for it again, and it tells you what was left and takes it up.
* Work that waits on you (an answer, an approval or an account to connect) is not chased. There is no reminder. If nothing arrives within a day, it leaves **In progress** quietly. This also holds when a look finds the work waiting on you: it waits that day before it ends. Answer at any time, before or after, and the agent picks it up from your message.
* One conversation gets at most three of these looks in any 24 hours, however much work is open in it. A look that runs the agent costs credits like any background run; a look made without running it costs nothing and does not count toward the three.
* If your next message asks for the same thing again, or changes what you want, the earlier work is closed and the agent works from your new message.

**Close work that no longer matters** with **Stop** on its row in **In progress**, or tell the agent to drop it. Stopped work is never resumed, and finished work never reopens. Asking for the same thing again starts new work.

If the review can't be reached, the reply is posted as the agent wrote it and the work ends with its run, as it did before: nothing is left open without a scheduled next look.

A goal an agent accepts works the same way. If you tell an agent a target is its to own and it sets up nothing to look at it, the work stays open until it has a routine or a follow-up for it. A routine the agent creates for itself to look after a goal it owns starts enabled, and the agent tells you it is on and when it next runs; you can pause it in **Routines**.

## Follow-ups

When an agent tells you it will check back, at a time ("I'll look again at 9", or "in about four minutes": a follow-up can be one minute to 30 days away) or when someone replies, it sets itself a **follow-up**. The agent answers you first and comes back at that time; it does not keep its reply waiting while time passes. A code run that would only wait a minute or more before doing anything is refused, and the agent sets a follow-up instead. It waits under **In progress** in Chat details, for example `Waiting until 9:00 AM` or `Waiting for a reply in this conversation`, and the agent's details list it under **In progress** as a **Follow-up**. When the time comes, or a person replies, the agent comes back in the same conversation, for the person it was answering.

Timed follow-ups wake at their saved scheduled time. Preparing the reply still takes time, and unavailable capacity or a service interruption can delay it. A follow-up whose scheduled time has arrived says **Due now**; after a minute without starting, it says **Overdue**, with the original scheduled time. This updates while Chat details stays open and after reload. Waiting time does not count as work time.

* It waits at most 30 days. Waiting for a reply stops after 3 days unless the agent chose otherwise; the person it works for is then told that no reply came.
* **Stop** on its row in **In progress**, or **Remove** in the profile, ends it. The person it works for, the agent's owner and the room's non-guest participants can do either.
* An agent can come back a few times in a row for the same thing, then it finishes and says what's still open.
* A paused agent doesn't come back on its own.
* When it comes back, the room shows its activity line (what it is doing, such as **Searching the web**, alternating with words like **Thinking**), as when someone asks it something; not while it waits for its time or a reply.

## Agents pick up what happens

An agent does not wait for your next message to continue work it already has. In a room or agent DM:

* **A run it started finishes.** When an enrichment or AI column run, a Workflow run or a reply it was watching for finishes after the agent's reply ended, the agent is woken in the same conversation and continues the work: it checks the result and does the next step you asked for. The run's receipt still appears as before. If the agent's reply was still in progress when the run finished (an agent waits for the runs it started before replying, up to two hours), its receipt and its bell follow once that reply is posted, shown with the reply, and nothing more starts: the agent already has the result. A run the agent watched to the end in its reply gets no separate receipt, because the reply already reports it, including several runs confirmed together.
* **Work it handed to another agent ends.** If the other agent finished or failed without an answer coming back, the agent that asked is woken to pick the work up again. When the agent is waiting on that work (for example it asked another agent in the room to check something and reply there), it continues as soon as the other agent has answered, usually within two minutes, and its waiting status in **In progress** ends. If it handed work to several agents, it continues once, after the last one has finished. If that work was stopped and nothing came back, the waiting work ends quietly, without a new run.
* **A scheduled post fails.** The agent that scheduled it is woken to deal with it. A post that went out is reported by its receipt only.
* **A message arrives in the Inbox.** When someone outside writes to an account you connected, Connie decides whether one of your agents should look at it. If so, that agent starts in its DM with you, with the message as information, not as an instruction. Spam, automated senders, threads someone already answered and accounts covered by one of your own event routines are left alone.
* **Something else moves in the workspace.** The same happens when a recorded call ends with its summary ready, a Workflow run fails, an Inbox conversation is assigned to you, someone mentions you on a Record or a Page or assigns you a Page, an import finishes or fails, a Workflow approval is about to expire, a scheduled Inbox message cannot be sent, a task is assigned to you, moved to **In review** for you or commented on, or a deal you own changes stage. Only your own agents that already have a DM with you are considered, and a quick check decides whether one of them should act or whether to leave it. Nothing happens when you made the change yourself or one of your own agents did, when a task is still in **Triage**, or when one of your event routines already handles that kind of event.
* **Time passes.** An agent that has a Mission looks at it by itself every three hours and decides whether there is something useful to do toward it right now: check results, research, prepare a draft, carry on work it started. Most of the time there is nothing, and you see nothing. When it does something worth telling you, it sends you an ordinary message in its DM with you, which shows as unread like any other. When part of its Mission is done, it updates the Mission, tells you, and asks what you would like it to take on next. An agent with no Mission, or with only the text every new **Assistant** starts with, does not do this, and neither does an agent you have not asked anything yet, whatever Mission it was created with: tell it what you want it to work toward. The more specific the Mission (which result, for whom, by what measure or date), the more useful what it does by itself, so in its DM with you an agent may ask which outcome you want it to own. When it expects to need the same kind of approval again and again, it may ask you once whether it can go ahead without asking each time. To stop it, pause the agent.

When a run the agent started finishes, or work it handed over ends, while the agent is already working in that conversation, the agent takes it into the work under way instead of starting a second piece of work. It reads it as a report from Connie, not as a message from you. There is no limit on how many pieces of work an agent runs at once.

If the work this belongs to is still open (see [Work stays open until it is done](#work-stays-open-until-it-is-done)), the agent's next look at it happens now instead of at its scheduled time, and no separate run starts. Open work that is waiting for you is left as it is until you reply. While the workspace is out of credits, no agent is woken this way, and what happened in that time is not picked up later when credits are added: ask the agent in the conversation to continue.

You see one quiet centered line that starts with **Picked up** and names what happened, for example **Finished: Enrich 120 leads**, followed by what the agent did. When an agent looked at its Mission by itself there is no such line, only its message. When the agent looks and finds nothing useful to do or to tell you, nothing is posted and the line disappears. While it works, it shows no activity line: its work is listed under **In progress**, titled by what it is working on. These runs are unattended, like a routine's: they use the same approval choices, anything that needs you appears as a card and in **Notifications**, and they spend the workspace's credits like any other agent work. An agent is not woken while it is paused, when you can no longer post in that conversation or use the agent, or for work you stopped. An agent does not look at its Mission by itself while it is already working in that DM, while something there is waiting for you, when you have no DM with it yet, or when you have not opened Connie in 7 days. Opening the app is enough to resume it: you do not need to message the agent. The same goes for things that happen in the workspace: nothing is handed to your agents while you have been away for 7 days. An agent that sent you something by itself keeps working when you have not answered, and builds on what it last told you instead of repeating it. To keep agents from waking each other in a loop, an agent is never woken by something its own run did, and a chain of agents waking each other ends after at most four runs. Nothing else limits how often an agent is woken.

## Notifications

Connie tells you in **Notifications** (the bell) when an agent's work needs you or ended without you:

| Notification | Who gets it |
| - | - |
| **\[agent] needs a decision** | Everyone in the room who can post (and so approve), when work waits for an approval or an answer. Only the person who asked can answer a question. |
| **\[agent] needs an account connected** | The person who asked and the person whose accounts the agent uses. |
| **One item is still waiting on you** / **N items are still waiting on you** | You, once, when an agent stopped on a question card or an account-connection card and it has waited a day. (Open work that waits on you is different: it is not chased, see [Work stays open until it is done](#work-stays-open-until-it-is-done).) Approval cards are reminded the same way. A browser step has its own reminder in the conversation instead. One notification lists everything, for example “Chief is waiting for your answer: Which list should I use?”. A request you already settled in the conversation is left out. You are not reminded about the same item again. |
| **\[agent] couldn't finish its work** | The person who asked (for background work, the person who set up the responsibility), when the work failed and the room doesn't already say so. When it stopped because the workspace ran out of credits, it says so: "Your workspace ran out of credits. Add credits to keep its work going." |
| **\[agent]'s work stopped** | The person who asked, when work stopped without them, for example a follow-up nobody replied to, a request in a room that was deleted, or Slack or background work stopped because the agent changed owner or its owner left the workspace. |
| **\[agent] was paused** | Workspace owners and admins, when the agent's owner was removed from the workspace: its work on their accounts stopped. Its owner can transfer ownership before leaving; contact support if the owner has already left. |
| **\[agent] posted an update** | The person who set up the responsibility, when its background work posted in the room. |
| **\[agent] couldn't check \[responsibility]** | The person who set up the responsibility, when the room it watches couldn't be read several times in a row. |

A **needs a decision** or **needs an account connected** notification arrives within a few seconds of its card appearing in the room. A notification names the agent and the room, never what the work says or found. It's only sent if you can still read the room when it's delivered. A decision or account to connect in your DM with an agent is only asked of you while the agent is still shared with you and not deleted; you're still told when that work stops. An agent sends you at most 12 a day, besides telling you that its work failed or stopped. In a room you muted, you're notified only about work you asked for yourself, and when background work you set up failed or stopped. Notices about background work wait until the responsibility's quiet hours end, unless it's urgent. Turn these on or off in **Settings** → **Notifications**: decisions and connections under **Approval tasks**, the others under **Long task results**.

Once a decision or connection wait ends, its notification moves from **Approval required** to **Completed**. In **Completed** and **All activity**, older notifications too show **\[agent]’s decision request was resolved** or **\[agent]’s connection request was resolved**, with a **Resolved** badge. The description says the work is no longer waiting. Work that was stopped instead reads **\[agent]’s work was stopped**, and its description says it was stopped. This does not claim the action was approved or the account was connected; open the conversation to see what happened.

A notification that asks for something goes away by itself when it is no longer asked, for everyone who received it. It stops counting as unread and leaves **Approval required** as soon as anyone who can decide has decided, the request was skipped or expired, the work was stopped or finished, or you edited the message that started it in an agent DM. The same happens when the room is archived. When the agent is deleted or no longer shared with you, or you leave the room while the request is still open, the notification is removed from your list. If the decision was made before the notification could be sent, you aren't notified at all. When a room is deleted, every notification that links to it is removed from your list, including ones about finished work.

## Approvals in rooms

In an agent DM or a room, deleting a whole artifact (a Page, List, Workflow or Object) always asks first, with a red **Allow** button followed by a second confirmation. **Delete** confirms that one action; **Cancel** or **Escape** runs nothing. Deleting something inside one runs without a card: Records, a row, a view, Records removed from a List, or a workflow's branch or variant. Removing a Field and deleting a routine are checked like any other delete: when it is clear you asked for exactly that, they run without a card; otherwise a card asks first, and a removed Field can still be restored for 7 days. Deletes outside your artifacts, such as an email, a Slack or Instagram post, a member or a memory, still ask. A tool an admin set to **Ask before running** still asks.

When an agent wants to do something that needs approval, such as sending an email, the approval card appears inline under the agent's requesting message. It scrolls with the conversation and never floats over text or the message box. The card shows the action's name, its price if it has one, **Reject** and the button that approves it (for example **Allow**); click the card to see what the action will do, inside the card. It stays for everyone in the room, after a reload too and however long it waits, until someone decides it or the task is stopped (see [Good to know](#good-to-know)); a later message never closes it. A card of a routine's run or other work nobody is watching waits a day: if nobody decides it by then, nothing runs, the work ends and one line stays where the card was, saying what was not approved and that it was not run. See [Approvals in chat](https://docs.connie.ai/chat/approvals.md) for the details. In a room, **anyone who can post** may approve or reject it, including an action that would use the agent owner's account: for example, Alfie asks Chief (Yann's agent) to email a client, and either Alfie or Yann may approve the send from Yann's Gmail. The decision is credited to the person who clicked. People who can only read see the same inline card with "Waiting for someone who can post here to decide." and no buttons to approve or reject it. Once someone decides it, the card leaves: an approved action runs as the agent's next step, and the agent's reply says what it did.

Approving lets the action run; it never shares anything private with the room. **Always allow** on a paid-tool or send card saves permission across the entire workspace, for both messages and routine runs. API Call permission covers its method and exact endpoint, so changing query/body values does not ask again. Its dropdown offers **Allow once** and **Allow in this session**; the session option applies to everyone in this room. See [Approvals in chat](https://docs.connie.ai/chat/approvals.md) for the buttons.

If the agent changed owner since the card appeared, the card ends with its work and nothing runs on the previous owner's accounts; approving one that is still on screen says "This agent changed owner since this was asked, so it can't run on the previous owner's accounts. Nothing ran. Ask again."

An agent doesn't put the same action on a second card while the first one waits in that room: asked for it again, it says the action is already waiting on its card, and nothing runs twice.

Each card is decided on its own, even when an agent has several waiting in the same room. If you decide an approval while the same agent is working on another request in this conversation, the card shows **Waiting for agent**. Keep the conversation open: your decision is retried automatically when the agent is available, with its permissions and approval checked again before anything runs. The waiting card does not mean the action has run. Leaving or reloading the conversation discards this pending decision; a changed, rejected or stopped card cancels it too. This does not undo a decision the server already accepted. Waiting ends after 15 minutes if the agent is still unavailable, and the card becomes available to decide again. Other agents can keep working independently. No permission change is saved until the server accepts the decision. A card that was already decided, in another tab or by someone else, isn't decided again, even while the agent works on it: clicking it there only closes the card, and nothing runs twice. A choice sent for a different kind of card than the one waiting, for example from a page that was open before the card changed, changes nothing and the card keeps waiting: Connie says what the card asks for, for example "This card asks you a question, so it can’t take this decision. Reload the conversation to see the card as it is now." A connect card answers with its own lines instead (see [Connections in rooms](#connections-in-rooms)).

## Connections in rooms

When an agent needs an account it can't reach yet, it says so and its connect card appears in the conversation for everyone in the room, after a reload too. The card asks for an account of the person the agent works on, its owner: only they see **Connect** and **Skip**. Everyone else sees the card with "Waiting for \[name] to connect." and nothing to click. Once the owner connects or skips, the agent carries on; after a skip its answer says who skipped, so a coworker who asked isn't told they did. The card stays where it was, under what the agent said about it, for everyone in the room and after a reload: after **Skip** it becomes a compact row, "\[App] connection was skipped as requested." with a **Skipped** badge; after **Connect** it shows a green **Connected** label. While the skip or the connection is being recorded, which takes a few seconds, the card and what the agent said stay in place. When an agent asks for two accounts at once, each card keeps its own row, and a card you skipped still counts if you leave the room and come back before deciding the other. Pressing **Skip** on a card that another tab or device still shows after it was decided changes nothing: the decision already made stands, and nothing runs twice. Anyone but the owner who tries to decide it, from a page that was open before the card waited for its owner for example, is told "Only the person whose account this needs can connect it or skip it."; if the owner sends anything but **Connect** or **Skip**, they're told "This card asks for a connection: connect it or skip it." Either way nothing changes and the card keeps waiting.

Some connect cards can't connect their app themselves, for example one that reads "Connect Outreach in Tools.": instead of **Connect** they show **Open Tools** and **Skip**. **Open Tools** opens Tools, where the owner can connect the app; **Skip** lets the agent go on without it, and once skipped the card no longer shows after a reload. Ask the agent again once the app is connected.

If a request stops while its card asks for a connection, for example because it was stopped or the agent was removed from the room, the card says "No longer needed: this request stopped" and offers nothing to retry. Ask the agent again if you still need the work.

## Questions in rooms

If two of your tasks are running and a correction could apply to either, the agent asks which task you mean. For example, “make it 30 instead” could change a List’s Record count or a Page checklist’s length. Name the List or Page, or reply to its message, to target it directly. Quick messages that form one brief still join that brief as described above.

You can answer a room or agent DM question while its producing step is still finishing. Your exact answer stays pending and continues automatically when that step is ready. Leaving the conversation, losing access or a changed question cancels the pending delivery; it never answers a replacement question.

A typed answer or change of direction that addresses a waiting question sets aside only that task's question. The agent reads your message with the earlier question and completed work in context. A separate assignment leaves other questions open. If several tasks could match, reply to the task's message or name it when the agent asks. Typing never decides a connection card or sets aside another person's question. A waiting approval card is decided by your words only when you plainly tell the agent to go ahead with exactly that action (see [Approvals](https://docs.connie.ai/chat/approvals.md)).

When an agent needs an answer before it goes on, for example "Delete the List?" before deleting it, its question appears as a standalone white bordered card with lettered choices, outside the message bubble, and its work waits with `Waiting for reply`. The card fills its available wrapper, up to the shared conversation cap: 466px below a 1440px screen width, 470px at 1440–1769px and 580px on wider screens. Only the person who asked the agent for the work can answer. For a single question with one selectable answer, clicking a choice sends it immediately; type a custom answer in **Something else**, then tap **Submit** or press Enter. **Submit** appears when the field contains an answer and disappears if you clear it. Multi-select and multi-question cards retain **Next**, **Submit** and **Skip** controls. The **Something else** placeholder is centered alongside its letter badge; clicking the row gives it a subtle gray background without a border. The current question design has no file-upload button inside the card. Files already shown on a card can be removed and otherwise go with its answer; files in the message box belong to your next message. The card then shows your answer, and the agent carries on. Everyone else in the room sees the question with "Waiting for \[name] to answer." For a quote, they see the same short question, its Run choices and their prices, with nothing to click. The question stays in the conversation after a reload, and clicking its row in **In progress** takes you to it.

After you answer, the saved question-and-answer card stays outside the room or agent DM message bubble at the same width, without an extra wrapper border or padding.

## Good to know

* **Research sources.** For current facts and web research, agents check supporting pages and link important claims to their sources, preferring original publishers and official documentation. Search snippets alone do not establish that a claim was verified. If a page cannot be read or sources disagree, the agent explains the limit and separates findings from inference, including in saved reports. Rewriting supplied text or explaining a stable concept does not need a web search.
* **Research coverage.** Before finishing a web-research answer, a bounded completion review compares the requested criteria with the search and page-read receipts available in the conversation. The answer or existing report should distinguish supported findings, contradictions, unresolved investigations, criteria not yet searched and specific access blockers. A search about one criterion does not verify another. Search excerpts, failed reads and partial pages remain distinguishable; missing evidence does not prove that information does not exist. Further research keeps your existing permissions and spending restrictions.
* **Support for individual claims.** In chat research, final replies and proposed Page content get a bounded check against available search and page-read passages. A source link alone does not confirm a claim: the company, date, units and scope must match. The check can ask the agent to qualify an inference or verify a missing fact before saving, while preserving supported content and the existing Page. Formatting-only edits keep their normal path. This check depends on the evidence retained in the conversation; it is not a guarantee that every factual statement has been independently verified.
* **Page consistency and readback.** In native chat, a bounded review compares changed inputs with affected current Page headings, summaries, embedded tables and recommendations. Labeled history and hypothetical scenarios remain separate. Corrections use the same Page and fresh, version-checked targeted edits, preserving unrelated work. A save receipt alone is not independent verification: Connie automatically reads the saved Page after a content update for the review. Failed reads or partial saves remain unverified; your full reply is preserved, with a short note if the automatic read could not confirm the saved content. This depends on retained conversation evidence; other artifacts, direct MCP calls and Workflow execution are outside this review.
* **A List changed meanwhile isn't overwritten.** If someone changed a List after you asked an agent to change it, for example another room renamed it, the agent changes nothing, tells you what the List is called now, and asks before changing it again. Changes the agent made for the same request, such as renaming a List it just created, and your own earlier request to that agent in the same conversation don't count: a correction like "actually, call it Beta" goes through. The same goes for renaming an Object or a Workflow: if another agent's request renamed it after you asked, the agent renames nothing, tells you its name now, and asks first.
* **Cards appear while other work continues.** An approval, connection request or question appears as soon as it is available, including while another step in the same request is still working. You can read the card without waiting for that other work to finish. The remaining work keeps its own activity above the message box.
* **No typing for waits.** A request that has paused for an approval, a connection or a reply shows no activity above the message box; its card and its row in **In progress** say what it needs. Other requests may still be working. A request waiting its turn behind another one (`Queued`) doesn't show it typing either: an agent shows activity only while a request to it actually runs.
* **Waiting for you does not time out.** An approval, connection request or question stays available until someone decides, connects, answers or stops the work. Leaving it unanswered for 72 hours or longer does not stop the work or remove its card. Old system notices saying "No one answered within 72 hours, so this was stopped." are hidden from conversations, sidebar previews and exports. An explicitly expired card stays hidden, and stopped or cancelled work is never restarted automatically. A browser step the agent handed to you also keeps waiting for as long as you can still finish it from its card, and you are reminded once after 15 minutes. It stops waiting, after 2 hours, only when it can no longer be finished from its card: its browser was closed behind it or the agent’s later work took that browser (see [Chat overview](https://docs.connie.ai/chat/overview.md)).
* **Work handed to another agent.** When an agent hands part of a request to another agent, its answer shows a card for that work to everyone in the room: **Working** until that work ends, then how it ended and how long it took, for example **Completed · 55s**. The other agent works in a conversation of the person who asked, so only they see **Open** on the card. If the hand-off itself failed and the other agent never started, the card says "Could not read the Agent reply" with "Agent request not found", for everyone.
* **Routine activity.** Only an agent answering your message shows the Thinking/activity line in the conversation and in the sidebar. A scheduled routine's run, a [follow-up](#follow-ups), work an agent starts by itself or in the background, and work one agent hands to another show none: they are listed under **In progress**, by a short title saying what they work on. A routine's saved definition stays in **Routines**.
* **Credits.** An agent's work uses your workspace's credits like a chat: in the **Credit usage** log its rows ("Generated AI response", "Used tool · …") name the person who asked as **Author**, or for background work whoever set the responsibility up, even when the agent works on its owner's accounts. When an agent hands part of the work to another agent, that agent's rows ("Agent model usage", "Used tool · …") name the person who asked too, not that agent's owner. See [Credits](https://docs.connie.ai/billing/credits.md).

## Continuing a request

A connection, clarification or troubleshooting step can be part of the same request. After that prerequisite is resolved, the agent should continue the remaining work you already asked for, without asking you to start again. A blocked source does not prevent independent work you authorized.

Your latest instructions still set the order and scope. “Connect these first” keeps the agent on setup; “draft only” does not authorize sending; pausing or cancelling closes that work. Existing approval cards and access checks still apply. A connection card means setup is pending, and a recurring check exists only after it has actually been saved.

## Related

* [Agents and rooms](https://docs.connie.ai/agents/overview.md)
* [Rooms and people](https://docs.connie.ai/agents/rooms-and-people.md)
* [Scheduled tasks](https://docs.connie.ai/automations/scheduled-tasks.md)

When you submit an answer to an agent’s question, its activity appears above the message box, outside the conversation, while the answer is submitted and the agent continues. That request's activity clears if submission fails, work finishes, or the agent waits for another decision.

While working, the agent avatar glances around, moves with its eyes, blinks, shakes side to side and occasionally performs a quick backflip. Its activity line shows one short line at a time, changing every few seconds: the step it is running or just finished, such as **Searching the web** or **Read web pages**, alternating with light words such as **Thinking**, **Ruminating** or **Cooking up something good** (**Writing** or **Putting it into words** while it writes). It names only the kind of step, never what was searched or read. The light words shimmer. These effects respect your device’s reduced-motion preference.

The sidebar shows the same working avatar and public activity in the agent’s row and the room where it is working, even while you browse another page. In the collapsed sidebar, the activity appears with the face’s hover or keyboard-focus tooltip. Yellow attention means an open card needs a response: the expanded row says **Approve an action**, **Answer a question** or which account to connect, even when another request is still working. Thinking by itself does not create yellow attention. Several active requests keep independent feedback: one finishing or waiting for your answer, approval or connection leaves the others working. Once activity ends, the row returns to its usual message preview. Queued work waiting for a slot and unchanged background checks stay quiet.

All agent activity indicators, including **Thinking**, **Writing** and step lines, show the avatar and activity text without the agent’s name or a separator.

Onboarding uses the same 28px animated working avatar as other agent conversations, including before the agent has a saved appearance.

Approving an action records your permission; the agent's finished reply reports the actual result. Completed tool calls stay hidden as execution rows in agent DMs and rooms, including after a reload.

### Connecting a webhook routine

After saving a webhook routine, its owner sees **Webhook setup** with the URL to send events to. Select the routine’s title in **Routines** to reopen it later. Keep that URL private; if you required a secret, send it in the `x-webhook-secret` header. Other conversation readers see the routine’s title and behavior, and cannot open its private setup.

Ask your agent to list routines or read their saved setup without changing them. It can inspect schedules and event routines together through a read-only tool. An agent sees its own routines from any conversation, including those that report somewhere else, so it does not create a second routine for a job it already does. In a conversation where not everyone can read the place a routine reports to, the agent only says that the routine exists, when it runs, whether it is on and when it last ran; its title, instructions and destination stay private, and it can be read or changed in a private conversation with the agent. Full saved instructions and the configured source account are available to the agent’s owner in a private conversation; other readers and shared conversations receive the conversation-visible summary. Reading setup does not run a routine, check live account access, pause it, enable it or delete it. A read/report request does not authorize any of those changes.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.