Share a public template
Choose Share agent in your agent’s details panel or its sidebar menu, then Share a copy instead. This opens your private conversation and puts Create a public template of yourself that I can share with somebody else. in the composer, ready to send. Any existing draft is kept below it. You can also ask this directly in the chat. Sharing is available to the agent’s owner while the agent is active. The agent gathers relevant instructions, reusable context and skills, then prepares an Unpublished template card. Choose View details, then use the Instructions, Context, Skills and Tools tabs to inspect what will be shared. Context and skill rows expand to show their full text, and the publish controls stay visible while you review. Ask the agent to change anything before publishing. Public-template cards have their own review layout: the agent’s name and publication status at the top, a large avatar on a tinted panel, a two-line description, and Publish and View details actions underneath. They use rounded pill actions in light and dark themes. Ordinary artifact cards keep their existing preview and icon/title/Open footer. Publishing returns a clean link such asconnie.ai/agent/<short-code>. Anyone with that link can view the published agent’s name, avatar, description and included content without signing in. View details opens the reviewed instructions, context, skills and tools. Add to workspace asks signed-out visitors to log in and returns them to the same shared agent. Choose a destination workspace to enter it. The agent review modal opens in the selected workspace without showing the template library. Review its context and integrations, then press Add agent to create an independent private copy. The workspace is fixed once selected; the review has no workspace-change or workspace-back control. Refreshing keeps the review in that workspace, and closing it returns to the workspace’s chat. Choosing a workspace alone creates nothing. Add agent opens your new agent’s private conversation directly. The review stays in an Opening agent… state until navigation finishes, and your new agent appears in the workspace sidebar. Visiting a link never creates or starts an agent automatically.
Links identify a published version. Publishing changes, or publishing again after Stop sharing, creates a new link and invalidates older ones. Copies already made remain independent. Older /shared/template/... links still use the same publication checks. Source profile names and private account information are not added to the landing page.
Choose Publish to create a link that anyone with the link can view and use. Copy link copies that link. Later edits remain private until you choose Publish changes. Stop sharing in the details view disables the link; existing recipient copies remain theirs.
A recipient creates an independent, private agent with the included skills. It uses their own connected accounts and their workspace’s tool permissions. Tool grants begin with approval required. It has no active routines, schedules or browser sessions from your agent. Your private conversations and account credentials are not part of the template. Essential private reference files must be adapted into shareable text or supplied separately by the recipient.
Create an agent
1
Open New agent
In the sidebar, click + next to Agents (New agent or room) and choose New agent. On a narrow screen, click Open navigation first, then + in the icon rail and New agent; the menu stays above the dimmed page.
2
Describe it
Give it a Name (for example “Chief of staff”). Pick a Color and a Face for its avatar. Faces are soft flat colored shapes with small character eyes: Flower face, Oval face, Dome face, Triangle face, Square face, Flame face and Hexagon face. At rest every face smiles; when its eyes are open, their pupils take the face’s color. The same face appears in the sidebar, messages, agent details and Slack.
3
Create it
Click Create agent. The new agent appears at the top of the sidebar; click it to start your DM with it. Agents and rooms then follow recent activity, so starting that first conversation keeps the new agent near the top instead of moving it below agents you have not opened. It starts without a mission. Tell it what you need in your DM. Its purpose is kept in AI context and has no profile field or editor.
Share an agent
Choose Share agent from your agent’s sidebar menu or Chat details. Invite people in the workspace, or choose Everyone in [workspace], then save. Sharing grants use of the same agent in each person’s own private DM; it creates no conversation until they explicitly open one. Only the owner can change sharing, including for Assistant. You can manage access while the agent is paused or its DM is read-only; doing so does not resume it. Guests cannot use agents. Sharing does not reveal anyone’s DMs or saved memory, or grant management controls. The agent still uses its owner’s permitted connected accounts and the workspace’s credits. Tool permissions, approvals, account exclusions and the audience’s access to Records and other resources still apply. The agent’s purpose and standing rules guide its work for everyone who uses it, so keep private facts in memory or your private conversation rather than in its shared instructions. Removing a person, or returning to invited people only, makes affected DMs read-only and blocks new messages and approval decisions. Saved history remains readable; work already started may finish. Agents already added to rooms can still answer for those rooms’ participants. Use Stop to stop ongoing work, or remove the agent from the room to stop its work there. Use a room when people should work in one shared conversation. Share a copy instead keeps the separate public-template flow: the agent prepares a reviewable copy, and nothing is sent or published until you choose to do so. Copies have their own owner and connected accounts, without the source agent’s conversations or saved memory.Whose accounts an agent uses
An agent works on its owner’s connected accounts. The person who creates an agent owns it. The profile says so: “Owned by you · works on your accounts”, or “Owned by [name] · works on [name]’s accounts”. The accounts follow the owner: an account the owner connects later is available to the agent too, and one the owner disconnects is gone, unless it was excluded in Manage accounts. A teammate who asks the agent for something in a room doesn’t lend it their own accounts. When an action would use the owner’s account, the agent asks first as usual; see Approvals in rooms.Who can talk to an agent
Because an agent works on its owner’s accounts, only the people it’s shared with can message it, put it in a room, mention it or give it work: its owner, the people it’s shared with by name, and everyone in the workspace when its owner shares it with everyone. A new agent is shared with nobody: only you see it and talk to it until you share it. In a room, everyone in the room can talk to the agents added to it there, without seeing your DM with it or your accounts. Your starter agent, Assistant, is shared with nobody: no one else can talk to it, workspace owners and admins included, unless you share it with them. In a room you add it to, the people in that room talk to it there. In your DM with a teammate’s agent, you can read and edit the pages made there, as in a room you can post in. Once the agent is no longer shared with you, or was deleted, you can’t edit them there any more. Every agent has one owner. Only that owner changes its name, appearance, mission, model, accounts, responsibilities or ownership, or pauses or deletes it. Workspace owners and admins have no extra control over someone else’s agent, including one shared with everyone. Sharing lets people use an agent; it does not let them change it. Historical edit shares continue to permit use without ownership controls. If an agent stops being shared with you, your DM with it stays readable but takes no new messages: in place of the message box it says “[agent] isn’t shared with you, so this conversation is read-only. Its history stays here; ask its owner to share it with you to keep going.” Its row stays in your sidebar, and clicking it opens that read-only history. When another agent messages it, your DM shows a small summary such as 6 messages with Blue and Connie QA bot where the exchange began. The same row updates as saved messages and replies arrive. Click an agent’s name to open the read-only conversation between the agents in a side panel beside your conversation. It replaces Chat details while open; your conversation and message box stay available, with your draft preserved. Each name opens only the two-agent thread with that agent. The panel’s header shows the two agents’ stacked avatars and a short AI-generated title based on their first saved message, without a name pill or exchange arrow. Message author names keep their agent colors. Naming loads separately, so you can read the messages immediately. Broadcasts use separate agent messages by default; a group room is created when you ask for a group. Click its Close agent conversation button or press Escape while inside it to close the panel; previously open Chat details return. You see only exchanges whose source and destination conversations you can still read; owning or sharing an agent does not give access to another person’s private conversation. Deleted, replaced and unfinished messages are not included in the count. Requested drafts and useful final outcomes also appear in the main conversation for your review. Routine acknowledgements stay in the transcript; after a broadcast, the coordinating agent reports a consolidated outcome.What an agent can use
Room, agent and routine tools run directly in chat. They cannot run inside a code script. If a script attempts one, that call is refused without changing anything or charging tool credits; unrelated script steps can continue. The agent can make the requested call directly in chat with the same targets and settings, subject to the usual permissions and approvals. An agent can use every tool in your workspace that the person it works for can reach. Tools are not chosen per agent: there is nothing to turn on or off for one agent, and a tool added to the workspace later is available to every agent straight away. Every agent also always has its own memory, its room helpers and the helpers that find tools and read skills; nobody can leave those out. What still decides whether a tool runs:- Workspace settings. A workspace owner or admin can turn a tool off for the whole workspace, or make it ask for approval, in Plugins. That applies to every agent and every chat. See Permissions.
- The owner’s accounts. An agent works on its owner’s connected accounts, minus the ones its owner excluded in Manage accounts.
- Approvals. A tool that asks first still asks first; see Approvals in rooms.
Manage accounts
Open the agent’s sidebar three-dot menu and choose Manage accounts, or click Manage accounts in its profile. It lists the owner’s connected accounts, one row per account, so two Gmail mailboxes are two rows. An account that is on is one the agent may use; turn it off to exclude it. When someone asks the agent to use an excluded account, by its address or its name, or asks for something only that account could do (for example sending from Gmail when its only Gmail mailbox is excluded), the agent says it isn’t allowed to use it and that it’s excluded in Manage accounts; it doesn’t read or send through another account instead, and it doesn’t ask you to connect the account again, because reconnecting an excluded account doesn’t allow it. An account the owner connects later is on automatically. Click Save to save; Cancel saves nothing. If the accounts changed somewhere else at the same time, it says “[agent]‘s accounts changed elsewhere. Open Manage accounts again to see the latest.” Only the owner can change them; anyone else can look, and sees “Only [owner] can change [agent]‘s accounts.”Change an agent’s model
Unless its owner chooses a model, an agent picks one for each request: Claude Opus 5.5 when you ask it to analyse, compare, recommend or work through a list item by item, and Claude Sonnet 5.5 for everything else, which costs about half as many credits. There is no model setting in the app: to change it, ask the agent, for example “Which model are you using?” or “Use Claude Opus 5.5 from now on.” It can list the Claude and GPT models it supports. A model you name is used for every request; “go back to the default” returns it to the automatic choice. Only the agent’s owner can change its model, by asking it directly in its DM or a room. Anyone else who asks is told so, and nothing changes. The change applies from your next message: the reply in progress, work already running and its pending approvals keep the model they started on. Routines and other work the agent starts by itself keep their own model, and voice calls are not affected. Changing the model changes no tools, permissions, memory or accounts. If the agent gets a new owner, it returns to the automatic choice until the new owner chooses.Ownership
An agent has one owner, shown in its details. To change who owns it, the current owner chooses Transfer ownership in the agent’s sidebar menu. The compact Transfer ownership dialog has a New owner dropdown, Cancel and Transfer ownership. Choose a member, then confirm; guests and the current owner are not offered. Cancel makes no change. The full profile does not open.The agent’s profile
Chat details, opened from the centered agent name, shows the cover, avatar, member count, In progress, Routines, Artifacts and Context. In progress hides when empty; saved recurring work stays in Routines. Its textured cover follows the agent’s avatar color, including custom colors, and updates when that color changes. The texture fills the cover without a colored outline. Rooms without a primary agent use a neutral cover. It does not show profile editing controls. The profile shows its owner, In progress, Routines, and Your rooms with [agent]. Only its owner can edit the name and appearance there. Mission stays in AI context and is changed by talking to the agent; there is no Mission or Instructions field, and agents have no managers. The agent’s owner also sees:
Pause agent and Resume agent are also in the agent’s sidebar options.
Voice
Every agent speaks with its own voice in an audio call. The profile shows it under the owner line, for example Voice · Hope. The agent’s owner can change it: click Voice · [name] to open a list of twelve voices (Hope, Mark, Alexandra, Dylan, Sarah, Eric, Jessica, Chris, Alice, George, Lily and River), each with a few words on how it sounds. Hover a voice, or move to it with the keyboard, and click its play button to hear a short sample; click it again to stop. Listening doesn’t change anything. Click a voice’s name to choose it: the choice is saved at once and the list closes. From the keyboard, the Up and Down arrows move between voices, Right moves to the play button, and Enter chooses or plays. The new voice is used from the next call the agent joins; a call it is already in keeps the voice it started with. An agent you never chose a voice for keeps the one it has always had. Two agents without a chosen voice never share one in the same call; if you choose the same voice for two agents, both use it. A duplicated agent starts with its own default voice, not its original’s. Everyone else sees the voice’s name without the list. If a sample can’t play you see “Could not play [voice]’s voice. Please try again.”, and a failed save shows “Could not change the voice. Please try again.” and keeps the previous voice.Work it runs on its own
Tell the agent what it is for in conversation; its owner can change that internal Mission the same way; saved recurring work appears in Routines in its details. An agent works when someone asks it. Work an agent was already given to do by itself, on a schedule or whenever something new is posted in a room, keeps running and shows in its details under Routines between runs, and under In progress while running. You can still ask an agent in its DM or a room to check something regularly; what it sets up appears in Routines too. When you ask from another agent’s DM or your own MCP client to set up recurring work for another of your agents, that work reports to your DM with that agent unless you name a reporting conversation. The DM opens for you if you haven’t talked to the agent yet. Work set up in a room the agent already participates in stays there by default. No new room is created for it; an agent makes a room only when you ask for one. Ask the agent “What scheduled tasks do you have?” to read its saved responsibilities, including their status and next check. To test one, ask the agent “Do a test run of the hourly task sweep.” It reads that responsibility’s saved instructions and carries them out once in the conversation. This manual test leaves its automatic schedule, status and next check unchanged. Normal tool approvals and credits apply, and the agent reports the result after doing the work. For another agent’s responsibility, it asks that agent and follows its task. In Chat details → Routines, recurring tasks appear with a generated title, a recurring schedule or plain event description, and an enable/disable switch in the agent’s color. Timed routines keep their time zone. Live executions appear separately in In progress. To test a scheduled routine or ask for its next run, ask the agent in the chat. To add one, click + beside Routines, which starts “Create a routine: ” in the message box, or just ask the agent; only an agent’s owner can assign it an event routine in a conversation they participate in. Its replies appear in that conversation. In the full agent profile, each responsibility shows what it is and where it posts, for example “Watches a room · Every 30 minutes · Posts in Leads”, when it next checks and what its last check found, for example “Next check 10:30 AM · Last: nothing new”. A room you’re not in is never named: it shows “Posts in a room you are not in”. Recent checks lists what each check found, including “Couldn’t check: …” when the room it watches couldn’t be read, so quiet work stays visible even when nothing was posted.
Only the agent’s owner can create and edit responsibilities. The non-guest participants of the room a responsibility posts in can also pause, resume or remove it. Everyone in that room can see it.
You can also tell the agent in the room, for example “stop posting the digest here” or “pause this until Monday”. It pauses that responsibility; for a pause until a time, it shows for example “Paused by Yann until Mon 2 Nov, 09:00 (Europe/Berlin)” and resumes on its own then, checking again from that time without making up the checks it missed. A pause can end at most 30 days ahead. An agent only posts in a room when someone asks it or one of its responsibilities or follow-ups reports there, so with none, it tells you there’s nothing to stop rather than promising to stay quiet. It knows which of its responsibilities and follow-ups post in the room you ask in; where it can’t change them from the conversation, it tells you which one posts there and that its owner or a non-guest participant of that room, can pause or remove it from its details, under Routines, rather than promising to.
What it posts
- Only something new. It posts when its work found something to report, and never the same result twice. A check that finds nothing posts nothing.
- A few times a day at most. An agent posts at most 6 updates a day in a room from its responsibilities.
- Quiet hours. Results wait until quiet hours end; then the latest one is posted. If by then someone can read the room who couldn’t read what the work found (the room got a public link, or someone joined who can’t read the room it watches), it posts Who can read this conversation changed while I was working, so I did not post that answer. Ask again and I will answer from what everyone here can see. instead. If the person who set it up can no longer read the room it watches, or that room was deleted, nothing is posted.
- Routine activity. A scheduled routine’s run shows no Thinking/activity line: only an agent answering your message does. Its definition stays under Routines; its active occurrence appears under In progress.
Credits and limits
A responsibility’s work uses your workspace’s credits like any request, with no separate limit per task, per agent or per day: it runs while your workspace can pay for it, and stops only when it can’t, as described in When you run out. A step with a known price, such as a priced tool, doesn’t start unless your balance covers it, and isn’t charged when it doesn’t; work already under way can take the balance a little below zero, as a chat reply can. When the workspace runs out, the work stops there: a run that can’t start ends with “Insufficient credits. Upgrade your plan to continue.”, and a run under way stops at the first paid step it can’t pay for, which isn’t run or charged. Either way nothing is posted in the room it reports to, and the person who set the responsibility up is told “[agent] couldn’t finish its work” with “Your workspace ran out of credits. Add credits to keep its work going.” (with a cancelled trial, the trial’s own message instead). It doesn’t pick up by itself once credits are added: the responsibility’s next run starts as usual. Its runs don’t count toward the usage limits of the person who set it up (how many messages they can send in a short time, and how much AI they can use in a day), so a busy responsibility never holds up their own chat, and their chatting never stops it. There is no limit on how many responsibilities an agent has or on how much it runs at once. A routine’s run, a follow-up or a later look at unfinished work starts whatever else the agent is doing, and cards nobody has decided yet never keep an agent’s routines from running. Work a responsibility hands to another agent uses the workspace’s credits the same way, with one difference: the other agent stops before any step whose estimated cost is more than what is left of the workspace’s balance as it was when that work started, saying “The workspace needs more credits for the next step. Add credits, then continue this conversation to resume the saved work.” So it can stop while the workspace still has some credits. A watch’s or review’s work starts from the new messages its check found in the room it watches: up to 20 a run for a watch and 12 for a review (as many as it asks about first). When up to twice that many were posted since its last run (40 messages, or 24 for a review), it takes the earliest first and its next run goes on from there, so none is left out. When more were posted, for example while the agent or the responsibility was paused, while its last run waited a long time, or in a busy room, one run takes the latest 20 (12 for a review) and is told how many were posted, which Recent checks shows too (for example “100 new messages”), and the next check looks only at what was posted after that. It only starts while everyone in the room it posts in can read the room it watches, the rule its checks follow. Otherwise it stops with the reason its checks show, for example “Couldn’t check: someone who reads the room it reports to cannot read the watched room.” If, while it works, someone who can’t read the room it watches comes to read the room it posts in (that room gets a public link, or such a person joins), it posts Who can read this conversation changed while I was working, so I did not post that answer. Ask again and I will answer from what everyone here can see. instead of its result.When something changes
Each paused routine shows why in Routines. Changes to the room or the agent take effect at once, not at the routine’s next run:- If the person who set it up can no longer post in the room it reports to, it pauses at its next run.
- If the agent is removed from the room it reports to, it pauses: “[agent] is no longer in the room it reports to.” An event routine there is switched off. Adding the agent back does not switch them on again; resume them when you want them back.
- If the agent is no longer shared with the person who set it up, it pauses: “[agent] is no longer shared with the person who set it up.” After a transfer of ownership, event routines its earlier owner set up are switched off, because they ran on that owner’s accounts.
- If the room it watches can’t be read three checks in a row, the person who set it up is told “[agent] couldn’t check [responsibility]”.
- While the room it reports to is archived, it is paused: “The room it reports to is archived.” This also happens when the last non-guest participant leaves the room. Unarchiving the room resumes the routines the archive paused, from their next time; one you had paused yourself stays paused.
- If the room it reports to or the agent is deleted, it ends: the routine is removed, an event routine’s trigger is disconnected, and work it had in progress stops, including anything waiting for an approval or a connection. History and completed actions stay.
- After a pause in service, it makes one check, not one for every time it missed.
- While its last run is still in progress (not started yet, or waiting for an approval, a connection or a reply), it doesn’t start another. A run that comes due meanwhile is skipped the same way, and Recent checks shows “Skipped: Its previous run is still in progress.” Once that run has ended, the next one runs as usual; a watch or a review then takes what was posted since its last run as Credits and limits says: up to 40 messages (24 for a review) 20 at a time (12 for a review), the earliest first, and more than that in one run with the latest 20 (12).
Delete an agent
Open the agent’s three-dot menu in the sidebar and choose Delete agent. Connie asks: “Delete [name]? [name] stops all its work and leaves every room it is in. Its messages, results and history stay where they are, so people can still read what it did. You can’t undo this.” After deleting, the agent leaves the sidebar, search and every agent picker, and its profile and your DM with it no longer open, for anyone: a link to either says it can’t be found. Nothing is erased: in rooms, its messages stay under its name so people can still read what it did. Under each request it was working on, the room says “[name] was deleted, so it stopped working on this.”; where no room can say it (a request from Slack, or one in a deleted room or a DM), the person who asked gets a [name]‘s work stopped notification instead. Its responsibilities end at the same time and make no further checks. If deleting stops part-way, for example on a connection error, deleting it again finishes the job. Deleting an agent never deletes a room, and deleting or leaving a room never deletes an agent.Good to know
- Taking part is not managing. Being in a room with an agent, or being able to message it, doesn’t let you change its mission, permissions, responsibilities or work.
- Asking an agent to manage agents uses your rights. When you ask an agent, in your DM or a room, to list agents, read their settings, or pause, resume or change one, it does so as you, not as its owner: it lists only the agents you can see, and it can’t change an agent you don’t manage. The same goes for an agent it hands your request to, whoever owns that agent: it too does these as you, not as its owner. In a room with other people, it names only agents everyone there may know of. An agent is deleted only by a person, from its menu in the sidebar (open its three-dot menu and choose Delete agent), never from a conversation.
- Your starter agent is an agent like any other: rename it by clicking its name in its details, choose its accounts in Manage accounts. If you delete it, it doesn’t come back.
- Old agents. When agents and rooms were turned on for your workspace, agents people had talked to became agents here, with their conversations: each person’s most recent private conversation with an agent became their DM with it, and their other conversations with it became rooms. Work that Connie or another agent handed to an agent for you isn’t a conversation you had with it: it doesn’t become a DM or a room, the people who could open it still can, and a link to it opens it read-only.