In February 2025 a Wyoming court sanctioned three lawyers for a motion citing nine cases, eight of them invented by the firm's AI tool. The draft took an afternoon; the cleanup took the case. A judge reads briefs, so it surfaced in weeks. Most software has no judge.
A generic `query(sql)` MCP server and a wide-scope on-call agent will out-confidently misdiagnose more pages than a narrow diagnostic API and a single-job harness, no matter how good the model behind them gets. Tool design and agent scope are the surfaces where the floor on hallucination actually moves.
A `DELETE FROM events WHERE created_at < NOW() - INTERVAL '180 days'` ships, and the disk does not get smaller. Partition rotation, covering indexes, and gp2's burst-credit accounting are what actually decide whether the storage dashboard is telling you the truth.
work_mem, shared_buffers, innodb_buffer_pool_size: three knobs, and the one that caused your OOM is the one nobody resized. A backend gets SIGKILLed, the postmaster can't trust shared memory, and the whole database runs crash recovery.
Forty analysts with a dashboard auto-refreshing, twelve microservices each polling a status badge, a connection pool sized four times past the core count. None of it trips the slow-query log, and all of it lands on the same cores.
METR's controlled study of sixteen experienced developers reported a 20% AI speedup; the stopwatch measured them 19% slower. The judgment that closes that gap is the deposit the agent now skips.
A runbook for software engineers who don't troubleshoot database performance often. What to check first, what the output means, and how to narrow down the cause before you act on it.
A team adds a third read replica because the primary is at 95% CPU. Lag gets worse. The slow-query log is empty. The first stop is `pg_stat_statements`, not the cloud console. This is the framing for the next five posts in the series.
An MCP server that exposes query(sql) against the prod replica is a wire, not a policy. The model writes whatever SQL the corruption floor produces, the rows it returns land in the LLM provider's trace storage, and the audit log shows one tool call from one service account with no prompt attached.
An AZ backup shipper that has sat in the toil backlog for two quarters lands in an afternoon. A diagnostic API in front of Redis the team always wanted but never staffed. The 10x claim collapses on the customer-facing PR, holds on the throwaway tool.