Backend engineer tests LLM agents for coding and review

During one working week, a backend engineer used LLM agents for Rails backend development, code review, specifications and domain research. Logs covered five days, one engineer and one large Rails codebase, with 2200 session files across four tools. Claude Code produced 730 session files, but only 80 had the engineer typing. The engineer typed 350 prompts and kept sessions live for 54 wall-clock hours, with two or more sessions running for 63% of that time.

Code review dominated. The engineer reviewed 19 merge requests, ran 16 four-tool panel runs, approved three merge requests and posted 44 line-anchored comments. Agents found real issues including a rounding error that lost a penny, a migration missing a statement-timeout guard, an unrelated one-time-password bypass change, a lookup that could hide valid records and tests that could not fail. The per-hunk reviewer was noisy, producing 59 candidate findings across three runs; 17 were confirmed and 42 were refuted, a 71% refutation rate.

The strongest lesson was that more agents did not automatically mean higher reliability. Agreement sometimes reflected stale worktrees, shared prompts or missed context, while divergent checks uncovered meaningful defects. Mutation analysis performed best in the logged week, finding 4 real issues from 4 findings by deliberately breaking code and observing test failures. The workflow still required human judgment: nothing was posted to colleagues without review, and AI expanded what was possible but demanded constant constraint-setting, verification and attention management.