The orchestrator and verifier patterns

Two patterns recur across the sources. The orchestrator pattern: a parent coordinates specialists in sequence, for example Planner (analyze requirements, make a plan) then Implementer (build it) then Verifier (confirm it matches). The "Everything About Codex" guide describes Codex's subagent manager decomposing work and dispatching parallel workers, each with its own context, which is how it claims to compress weeks of work into days by running independent threads at once. The likely near future it sketches: an architect agent proposes, builder agents implement in parallel, a test agent verifies, a reviewer critiques, and a human signs off.

The verification agent deserves special attention because it fixes a real, common failure. AI agents love to declare victory. They mark a task done when the implementation is partial or broken. A verifier is a deliberately skeptical subagent whose only job is to check claims: identify what was claimed, confirm the implementation exists and runs, run the tests, look for missed edge cases, and report what actually passed versus what is broken. Cursor's example prompt instructs it to "not accept claims at face value. Test everything." This is the multi-agent expression of Chapter 10's "an agent cannot mark its own homework," and it leads straight into Chapter 13's security agents.