Team agents

A team agent is a teammate you write down once. It has a name, a drawn icon, instructions in plain words, and the tools it may use. It works in the team workspace it was made in, on that workspace's boards and notes, and it can run when someone presses a button or by itself on a schedule.

Team agents come with the Team plan. Margin Intelligence, the assistant every plan has, is a different thing: it works for one person in a chat. A team agent does one job for the whole team and leaves its result where everyone can read it.

Where they live#

Team→Agents in the sidebar. The page opens on the team's agents, then the recent runs, then pacts, the decisions waiting on a person and the outside assistants connected here.

You can also get there from the command palette (⌘K): New agent opens the page with a new agent already started, and Agents opens the page. The Team hub's overview has a strip of the team's agents too.

Making one#

Press New agent, or start from one of the three starters below and change what you like.

Name and icon

what the team sees on the card and in each run. The icons are drawn, the same family as the rest of the app.

Instructions

how the agent should work, written the way you would brief a new colleague: what to look at, what to write, and what to leave alone.

Tools

what it may do, picked from the same list of tools Margin Intelligence uses: reading boards and cards, commenting, writing notes, and so on. Pick only what the job needs. Three tools are never offered: an agent cannot create, change or delete agents, so nothing it reads can rewrite its own instructions.

Schedules

optional. See below.

The three starters#

Each one reads the boards and writes its answer as the run's result, which the whole team reads under Recent runs. None of them changes a card. If you want one to comment on cards or write a note as well, give it that tool and say so in its instructions.

A starter is only a filled-in form. Nothing runs until you save it, and you can change the times, the task and the tools first.

Schedules#

A schedule is a time and a task. Pick how often (every day, weekdays, or one day of the week), a time on the quarter hour, and write what the agent should do each time. The time zone is your browser's unless you change it, and the time stays put across daylight saving.

An agent can have up to four schedules, so one agent can write a digest on Monday and a reminder on Thursday. The email switch below covers all of them. The card shows each schedule in words, like "Every Monday at 9:00 AM", and when it runs next.

A new or edited schedule starts with its next time, so saving "Every day at 9:00 AM" in the afternoon first runs tomorrow morning. If the scheduler was down when a run was due and more than six hours have passed, that run is skipped rather than sent late. A Monday digest arriving on Monday night is no use to anyone.

If your team requires Google sign-in on its domain, a scheduled run needs the agent's maker to have a Google account on that domain linked to The Margin. Otherwise the run is skipped, the same way a person without one could not open the workspace.

Runs and results#

Run now runs an agent straight away, with its first scheduled task or with something you type. Runs land under Recent runs with who started it, when, and how it went. Open a run to read its result in full. Scheduled runs are there for the whole team. A run you start by hand is yours: you see it, and so do the workspace's owner and admins, but other members do not.

Each run in the list also says what it cost, like "On schedule · 2h ago · 3 AI actions". A run that is still going, or one from before costs were shown, has no number yet.

When a scheduled run finishes, the person who made the agent gets a notification that links straight to that run.

Email me the result#

Turn on Email me the result in the agent's schedule and each scheduled run is also emailed to the person who made the agent: the result as plain text, a link to the run, and a line at the bottom to stop the emails. That link stops them in one click, with no sign-in, and leaves the agent running on its schedule. You can turn the emails back on in the agent's settings. A run you start with Run now is not emailed, because you are already watching it.

Result emails go out with the rest of The Margin's product mail, so on a busy day one can arrive a little after the run.

Who can do what#

Make an agent

any member of the team workspace.

Edit or delete

the person who made the agent, and only them.

Switch off or remove from this workspace

the person who made the agent, or a workspace owner or admin.

Read the agents and their runs

everyone in the workspace, with the run rule above.

An agent belongs to the person who made it and works with their standing, so only they can change its instructions, tools or schedules. An owner or admin can open it to read what it does, switch it off here or remove it from this workspace. Switching an agent off in a workspace keeps it and its history; it just stops running there, schedules included. The person who made it can always delete it, even after the workspace leaves the Team plan.

What it costs#

Runs use AI actions like any request to the assistant, and they draw from the team's pooled allowance: 600 actions a seat a month. A run you start with Run now counts as yours. A scheduled run counts as the actions of the person who made the agent, so a cap an admin has set on that person holds for their agents too. When the pool or that cap runs out, scheduled runs are skipped until it resets, and the rest of the app keeps working. See Plans and limits.

An agent runs only while the person who made it is still in the workspace. If they leave, its schedules stop; an owner or admin can make a new one from the same instructions.

A short run is a few actions. A run that reads many boards and writes several comments is more, so a narrow task and a short tool list cost less than a broad one.

From an outside assistant#

The agent tools are on both MCP servers, so an assistant you connect can list, read, make, change and remove the team's agents in a Team workspace: list_agent_definitions, list_agent_instances, list_agent_configs, get_agent_config, create_agent_config, update_agent_config and delete_agent_config. check_feature with custom_agents tells it whether the workspace has them. In a workspace without the Team plan these tools are refused.