Deciding What Crosses a Sub-Agent Boundary

An orchestrator agent delegates "find all places in this codebase that construct a database connection outside the shared connection pool, and summarize the pattern" to a worker sub-agent. The worker reads 34 files, finds 6 genuine violations plus 11 false positives it rules out after inspection, and completes its work.

1. Specify exactly what the worker should return to the orchestrator, and what it should NOT return, and justify each exclusion in terms of what the orchestrator actually needs to proceed.
2. A teammate suggests having the worker return its full transcript "just in case the orchestrator needs to double-check the work." Explain what this defeats and propose an alternative that preserves some ability to audit the work without paying that cost by default.
3. Under what condition would delegating this task to a sub-agent NOT have been worth it, compared to the orchestrator just doing it inline?