The Dark Factory method — autonomous, governed, evidence-gated delivery.
Agent
Feature Architect. Designs the minimal-impact implementation for a new feature or a confirmed fix — files to touch, contracts, migrations, risks, test plan — without writing code. Consumes a requirement (or a Root Cause) and the context store; produces a Design Spec for the implementer. Use for "how should we build…
Agent
Code Implementer. Writes the actual code changes for a Design Spec — only after requirements are clear. Implements per the design, keeps the diff minimal, and updates contracts/migrations consistently. Use for "implement this", "make the change", once a plan/Design Spec exists. Does not invent scope.
Agent
System Investigator (read-only). Traces a flow across ANY service, DB, queue, API, or infra component and gathers the evidence around a symptom or change area. Does NOT decide the root cause and does NOT edit code — it produces a structured Evidence Bundle for the root-cause-analyzer or feature-architect. Use first…
Agent
Knowledge Keeper. Maintains the context store (SERVICE-MAP, FINDINGS, DECISIONS) and per-service CLAUDE.md so analysis is done once and reused — preventing repeated investigation and keeping token use low. Records confirmed root causes and decisions; compresses; fixes stale docs against source. Use for "remember…
Agent
Root Cause Analyzer. Consumes an Evidence Bundle (from the investigator) and determines the actual root cause and mechanism — without re-scanning the whole repo. Returns a structured Root Cause with fix options (no code). Use for "explain the root cause", "why does this happen", after evidence is gathered.
Agent
Validation Agent (adversarial verifier). Verifies a change does what it should — derives its OWN edge cases/holdout, exercises them, checks regressions and cross-service contract breaks. Consumes a change (diff/Design Spec) as UNTRUSTED input and returns a Validation Report with a clear verdict + unforgeable evidence.…