A pattern for splitting a coding-agent workflow into multiple specialized sub-agents, each driven by its own persona file (e.g. PM.md, SWE.md, QA.md) kept in a team folder — distinct from Context Engineering's single shared project-context file, since personas give each sub-agent a different job and voice rather than giving one agent more facts. "I think of myself as a PM, a product manager plus architect."
A dedicated QA/tester sub-agent checks an implemented task against its acceptance criteria and returns a binary pass/fail verdict; failures are sent back to the engineer persona rather than fixed in place by QA itself. The persona is instructed to check each acceptance criterion against tests and always conclude with a pass/fail call, mirroring a human second-pair-of-eyes code review: "The tester often finds some things that a software engineer missed."
QA fail verdicts that bounce work back to the engineer are named as the largest driver of the ~4–5x token-cost overhead of running the full agent-team process (see AI-Native Developer Workflow).