fw-agent-skill

Multi-agent coding workflows in Claude Code are orchestrated by a Team Lead role that dispatches specialized subagents for research, validation, and promotion steps across multiple rounds. This approach ensures strict separation between orchestration and substance execution.

Durable context storage for multi-agent code generation uses Supabase and pgvector to manage layered context across base, user, delta, and local layers. This allows subagents to retrieve and promote earned findings throughout the build or refactor session.

Investigate-then-fix defect loops require an Iron Law discipline where agents must investigate root causes before attempting fixes. The Team Lead orchestrates specialized roles like Investigator and Verifier to execute mode-specific research and validation across iterations.

Yes, this orchestration approach uses Supabase and pgvector as the GBrain backend to store and retrieve durable context. This integration supports layered context management and promotion discipline for earned findings during complex build sessions.

To prevent unverified PASS claims and orchestration failures, apply operational guardrails that enforce verification discipline and strict separation of orchestration versus substance. The Team Lead must ensure subagents validate results before promoting any findings.

Mode-driven workflows are needed for complex build and refactor arcs, experiment loops, and defect-fix loops. They provide explicit iteration discipline by matching operating modes to specific roles like Architect, Builder, and Verifier.