AI can rewrite the query and propose indexes, but it needs real database access to prove the fix works.

N+1 query optimization follows recognizable patterns (eager loading, JOIN rewrites, batch fetching), so the structural approach is repeatable. However, the right fix depends on the specific schema, data distribution, and ORM in use, requiring per-instance judgment.

Success criteria are unusually crisp: query must run under 20ms, and before/after benchmarks are required. The agent has a clear, measurable finish line, which is favorable for automation.

The agent needs read access to the Flask codebase, the database schema, a live or representative database instance to profile against, and permission to run EXPLAIN/ANALYZE and timing benchmarks. Most production setups don't grant this by default, making this the primary automation blocker.

A bad rewrite could introduce incorrect query results (wrong JOINs silently dropping or duplicating rows) or a poorly chosen index could degrade write performance. These are reversible with version control and migration rollback, but data correctness bugs can be subtle and costly if they reach production.

Choosing between JOIN strategies, deciding which indexes are worth the write overhead, and validating that the rewritten query returns semantically identical results all benefit from human review. The agent can do the heavy lifting, but a developer should verify correctness before merging.