Skip to content

Coworkers ​

An agent can hand work to other agents: its coworkers. It asks one of them a question, gets the answer back, and carries on.

Not the same as a team of people

An organization's teams are groups of people you share things with. Coworkers are agents calling agents.

When it helps ​

Delegation is worth it when the work splits into jobs that need different instructions, tools or models: a manager that plans and writes, a researcher that reads the web, a specialist that knows one system well. It is not worth it for "do the same thing twice": that costs two model calls for one result.

Setting it up ​

Pick its Coworkers when you create an agent, or later under Coworkers on its Configure tab. Then:

  • Its model must support tools: handing work over is a tool call.
  • Coworkers must be published, and belong to the same owner. Work never crosses from one owner to another.
  • It goes one way: the agent can ask its coworkers, not the other way round, unless you add it to theirs too. A chain can be three deep, and loops are refused: an agent cannot end up calling itself.
  • A coworker follows its latest published version, so improving it improves every agent that works with it.

What the person sees ​

In chat, a delegation appears as a card: which agent was asked, what it was asked, and what it answered. Nested delegation is indented. Nothing happens off-screen.

What it costs ​

A delegated answer is a full run of the other agent: its own model calls, its own tools, its own tokens. Budgets apply at each step, and the whole chain shares one limit on steps and tokens, so confused coworkers stop rather than spiralling: at the limit, each agent answers with what it has, and the reply ends with "Cut short", so you can ask it to continue.

Watch the trace the first few times. The usual surprise is an agent that hands work over far more often than you expected. The fix is in its instructions, saying plainly when to do the work itself.

CrewHall