orchestrator
splits work, checks blockers
Multi-agent terminal runtime
Run teams, not one agent at a time.
Loops turns project files into a live agent team with an orchestrator, coworkers, task updates, and clean handoffs.
loops --team --cwd .
ready
Team
4 activefrontend-dev
implements focused UI slices
qa-engineer
runs checks and acceptance notes
docs-specialist
keeps run context readable
Group channel
turn 07orchestrator Task snapshot loaded from tasks.md.
orchestrator frontend-dev owns the CLI chrome cleanup.
frontend-dev Working in packages/cli/src/tui.
qa-engineer Waiting for changed paths before checks.
Tasks
3/5 done- Project contract approved
- Provider models selected
- CLI chrome cleanup
- Website team runtime UI
- Final deploy check
orchestrator -> claudeCode
implementation -> codex
review -> opencode
orchestratorCoordinates work and decides when a run is complete.
coworkersImplement, review, document, and report focused slices.
handoffsTask updates and changed paths stay attached to the run.
Team
Every role has a job, a provider, and a next turn.
Orchestrate
The lead reads the contract, assigns task IDs, and asks only the questions that unblock the run.
Execute
Coworkers make one focused slice of progress, report files, and return control to the runtime.
Verify
Review and QA agents work from changed paths instead of scanning a mystery transcript.
Contract
The team starts from files, not vibes.
goal.md, plan.md, spec.md, tasks.md, and roles.json define what the run is allowed to do.
- Auto-enter team mode when the contract exists
- Save named teams under .loops/teams
- Override provider models per role
Handoffs
Runs end with state the next team can use.
Deliveries, blockers, task updates, acceptance notes, and changed paths survive after the terminal scrollback is gone.
CLI chrome cleaned
Team runtime UI active
Deploy verification queued