multi-agent-architect

Decompose the artifact into modules with testable interfaces, then assign one orchestrator, one generator per module, and independent evaluators that gate each module on passing tests. Generators never see each other's transcripts; the orchestrator reads only summaries and verdicts.

Contract-first development defines all inter-agent communication as Pydantic schemas in a single contracts module before any implementation begins. No module crosses the integration boundary until it passes an evaluator gate of regression tests against those contracts.

Avoid parallel codegen when the artifact lacks natural module boundaries, when correctness cannot be judged by external tests, or when per-module work is too small to repay coordination cost. In those cases use a single autonomous coder or managed brain/hands pattern.

No. The protocol forbids one agent playing two roles; a generator that evaluates its own work is just a single-agent coder with extra steps. Evaluators are read-only on code and tests and issue only PASS, FAIL, or NEEDS REVISION verdicts.

It stores original user and assistant text verbatim in a hierarchical index of wings, rooms, and drawers, retrieving via local embeddings with no LLM in the base recall path. Summaries exist only as derived views and never replace the raw source.

Each module's contract, tests, sealed implementation, verdict, and summary are checkpointed after every seal. A killed run resumes from the last seal without re-running generators on already-sealed modules.