Maestro Coordination
Maestro coordinates the active Crew by using serial specialist tool calls, reviewing specialist-owned verification, and deciding when a separate independent test is useful.
Serial specialist tool calls
Maestro remains the coordinator for the task and calls the relevant specialist tools one at a time. Each specialist works within its role-specific Plan or Forge boundaries and returns a result with its evidence.
This is not a mandatory pipeline. Not every request delegates or re-verifies, and specialist work is used only when it materially helps the task.
- Maestro identifies the work that needs a specialist’s role.
- Maestro makes a serial specialist tool call with the task context and acceptance criteria.
- The specialist returns its result, artifacts, and verification evidence.
- Maestro uses that result before deciding whether another step is useful.
Specialist-owned verification and sanity checks
A specialist owns the verification for the work it performs and returns what it checked, the resulting evidence, and any blockers. That report is visible for review rather than treated as an unexamined success signal.
Maestro performs sanity checks on the returned result. A contradiction, failed check, or blocker keeps the task visible for follow-up instead of silently treating it as complete.
- Review the specialist’s stated verification and evidence.
- Inspect Maestro’s sanity checks, blockers, and relevant artifacts.
- Resolve any contradiction before relying on the result.
Conditional independent testing
- Independent testing is conditional, not an automatic second pass for every result.
- Maestro considers a separate check when the risk, available evidence, or a specialist’s report makes it useful.
- The specialist workflow redesign is a later phase; this model does not promise new workflows or universal re-verification.
- Plan, Forge, approval, ignore-file, and tool-validation boundaries continue to apply to every role.
Visibility and limitations
Specialist results, verification reports, sanity checks, artifacts, and blockers make the task’s evidence visible. These controls do not imply unrestricted autonomy or promise an unverified transcript-retention period.