Agents, delegation and team chat

You are drafting a page with your assistant when you ask it, in passing, what the three competitors charge. It keeps writing with you, a card appears saying it has briefed the researcher, and a few minutes later a second card arrives with the answer. The draft was never interrupted, and the research never landed in the middle of it.

That exchange is what this page is about: why your Player One is a team rather than one assistant, how work moves between them, and how to pick the right size of machinery.

Why a team beats one very capable assistant

An agent is more than a name and a picture. Each has a role, its own workspace, its own memory and a bounded set of tools, and the boundary is the point. Because the researcher only ever does research, what it learns about your sources, your standards and the shape of a report you actually like lands in one place, ready the next time you ask. A generalist doing a little of everything spreads the same learning across a hundred unrelated conversations, where none of it adds up.

The comparison with a human team holds further than you might expect, and then stops. Your agents specialise and get better at their own work, but everything they know about your business they know because it was written down or learned in the work with you.

Your installation arrives with a small team: a main assistant who takes the first ask, a general worker for routine side jobs, and a coordinator whose only job is to plan and hand out work. Specialists come from the gallery as you add them. They are on Manage > Agents (/agents), and Agents describes what one is made of.

When a new agent earns its place

Three signs are worth watching for, and you usually want at least two before creating anything.

• The work recurs. You keep asking for the same kind of thing, and keep giving the same background first.
• There is a distinct body of knowledge. Your refund policy, your house style, your suppliers: a subject in its own right that would clutter another agent's memory.
• You want a boundary. Something should touch your invoices but never your website, or answer customers without seeing anything you have not given it.

Creating one is a conversation rather than a form: you say what it should be good at and watch the card fill in beside the chat as your assistant builds it, with its own workspace and memory from day one. Choose a project first, because an agent belongs to the project you are in and New agent is disabled in All projects.

The field that matters most is the shortest. What the agent is good at is what the coordinator reads when it decides who gets a task, and matching is positive only: nothing records that an agent is a poor fit, so if you want it chosen for a kind of work its strengths must say so. Rewrite them under Skills on the agent's own page, opened from Manage > Agents (/agents).

Just as often the answer is not to create one. A second agent doing what an existing one already does is a memory split in half, and neither half compounds; a one-off piece of work needs a helper, not a hire. Grow your staff covers building one properly.

What you see when agents talk

When your assistant briefs another agent a card appears in your conversation naming who it went to, and when that agent answers a second card arrives with the reply. Both are attributed to the agent that sent them, not to you. Beside the message box sits a row of faces, one for each conversation this chat has spawned, with the chat it works for first: hover one to see its task and how long it has been working, and click it to read that helper's own chat in full. When helpers start helpers of their own and the team runs three levels deep, a map button at the end of the row opens the helpers map, with every chat laid out by level and You are here on the one you are in. The wider traffic between agents collects at Messages (/activity), reached from the notification bell or the dashboard.

How much of that traffic shows in your chats is one setting, Team messages, all chats, in a chat's more menu: Show draws every card, Compact folds each run of cards into one line you can tap open, and Hide leaves only a small marker with the count.

Two habits make this work. The helper cannot see your conversation, so its brief must stand on its own, and a thin answer usually means a thin brief. And if you take the work back, say so: a helper's instructions are frozen at the moment it started, so one that goes quiet and later wakes will happily finish an old plan on top of your new one.

Why side work belongs somewhere else

Everything in a conversation is read again on every turn, so a thread that has swallowed forty search results is slower and dearer for the rest of its life, and eventually gets compacted to stay workable. Sending the noisy part elsewhere and taking back a paragraph of findings keeps the thread short and on the thing you care about. Your own attention works the same way: you wanted a number, not a research log.

See Tokens, cost and caching for what a long thread costs and Conversations for what compaction does to one.

Choosing the shape of the work

Use the smallest shape that will finish the job, and move up the moment you find you picked wrong. Escalating is cheap; a half-finished result from too light a shape is not.

• Answer it here. A question, a lookup, an opinion. Nothing starts and nothing runs.
• Do it here. One clear deliverable you want to steer turn by turn. You stay in the loop.
• Hand a slice to a helper. The task stays here, but one heavy or noisy part goes to a background agent and comes back as a result.
• Start a job. Several parts, more than one skill set, or a scope that needs exploring first. A coordinator plans it, picks an agent per task and reports back while you do something else. See Jobs.
• Make it a process. The same work done the same way every time, with ordering, approvals and a record of what happened. Flows are early access. See Flows.
• Schedule it. Something simple that should happen every morning without you asking. See Scheduled tasks and watchers.

The line that catches people is between the last three: if one agent running one prompt covers it, a scheduled task is enough, and if it needs ordering, waiting on a person, or it runs constantly, a written process is cheaper than waking an agent every time. Give it work shows each in action.

Organising them into teams

Once you have more than a handful of agents, tell your assistant how they fit together: who leads, who mainly does what, who reports to whom. It writes that down as a team page that every internal agent reads, so work arriving at the wrong desk gets passed to the right one without you. A team also learns standing instructions the way a skill does, "always hand invoices to the bookkeeper", and when a team outgrows one page its manager gets a page of its own. There is no org-chart screen; the conversation is the control, and Agents describes what the page holds. Customer-facing sandboxed agents stay outside teams by design.

Seeing who is working right now

Manage > Active Agents (/manage/active-agents/by-project) lists the conversations running this second, grouped by project, with each reply moving as it is written. Open chat takes you into one, and Stop agent interrupts a run without deleting the conversation or the work already done. A queue panel above the list explains anything that has not started: your installation runs a set number of agent turns at once and the rest wait with their position shown. Nothing queued is dropped, and chat turns never wait behind background work.

Agents that face your customers

An agent you point at the public is built the other way round. It is standalone: it reads its own instructions, its own knowledge and whatever you have shared with it, and nothing else. No shell, no keys, no general search, no way to widen its own permissions, and no place in your team. It never joins a job, a delegation chain or team chat, by design rather than oversight.

That separation is what makes it safe to hand a channel to strangers, and only you and your main assistant can change what it reaches. Create a support agent for your customers walks one through; Vault and sandboxing explains the boundary.