multiagent-orchestration

Decompose the work into units with explicit dependencies, then spawn one subagent per independent unit simultaneously. Units with no dependency between them run in parallel by default, and downstream units unblock as soon as their dependencies resolve.

It is a hard gate where a setup subagent starts and verifies every required service, database, and watcher before any coding lane begins. An unstartable service is reported as a blocker rather than silently proceeding against a degraded environment.

Use it when a request decomposes into independently executable units that would otherwise run one at a time, such as parallel feature lanes or concurrent test and implementation streams. Single-threaded tasks gain nothing from the orchestration overhead.

Yes, the plan-before-work gate is non-negotiable. The orchestrator presents a markdown plan with the agent roster, environment manifest, Mermaid execution map, and gap report, and waits for explicit approval before spawning any implementation lane.

The orchestrator runs a debug and retry loop for the failing unit. A persistently failing unit is escalated as a surfaced blocker rather than retried indefinitely, and material plan changes trigger a pause and re-approval.