PROJECT FIVE - MULTI-AGENT AND PRODUCTION HARDENING

EXAMPLES Example 1: Supervisor and subagents. A learner splits a harness into a supervisor that delegates to a researcher and a writer and confirms the combined system produces a better report than the single agent. In the lab the matcher and writer play those roles, and the key check is that the writer receives only the slim dictionary. Example 2: Cost and latency hardening through routing. A learner uses a cheap, fast model for simple classification subtasks and a more capable model for final synthesis, then measures cost and delay. In the lab the ledger measures cost per tier, and the arithmetic compares it with an all-strong run. Example 3: Reliability hardening. A learner adds backoff retries and a maximum-iteration safeguard, then deliberately triggers repeated tool failure to confirm the safeguard engages. That is the LAB_FAIL_TOOL test. Example 4: Deployment preparation. A learner writes a short production-readiness checklist for the capstone and honestly marks which items are and are not satisfied, treating gaps as the remaining work list. The checklist above is a model to copy.