Wiki / Use

Team mode#

Team mode lets one agent lead a small team that works at the same time: a project manager, a tech lead, researchers, developers, a tester and a reviewer.

Turning it on#

Either way adds the team agent to the tab cycle:

{ "team": { "enabled": true, "parallel": 3 } }

or, for the running session, /team on (and /team off). parallel is how many workers may run at once: 3 by default, 8 at most.

How a run goes#

You switch to the team agent and describe the goal. The orchestrator does not write code itself. It follows this order:

  1. Plan. It asks pm for scope, steps and done-criteria, and for code work techlead for a design and which files each developer owns.
  2. Do. It runs researchers on different questions and developers on different files together in one team call.
  3. Check. tester runs or writes the tests, reviewer reads the changes. Failures go back to a developer with the exact error.
  4. Verify. The orchestrator runs the build or tests itself before it says it is done. It never takes a worker's word for it.
  5. Report. What each role did, the verified result, and what is left.

While workers run, the transcript shows team ▸ developer: api started and team ■ developer: api done lines.

The roles#

RoleReadsEditsJob
pmyesnoscope, numbered steps, done-criteria, risks
techleadyesnodesign, files per developer, interfaces, the check to run
researcheryesnoanswers one question with file:line evidence
developeryesyesimplements the files it was assigned
testeryestests onlyruns the project's tests and reports each failure
revieweryesnolists problems with severity, file:line and a fix

Workers get fresh context and only the tool list their role needs.

The team tool#

The orchestrator calls one tool with up to 8 tasks:

{"tasks": [
  {"role": "researcher", "task": "Where is the config file parsed? Answer with file:line."},
  {"role": "developer", "name": "api", "task": "Add a --json flag in src/cli.rs ..."}
]}

A failing worker is reported as FAILED with its error and does not stop the others.

Permissions and cost#

Workers share your permission dialog and rules, so requests queue up; /yolo approves them all. Each worker is its own model conversation: a team run uses several times the tokens of a single agent. On a local model with a small context window, keep parallel low.

Edit this page on GitHub