The hardest agent bugs are the ones where the output looks correct
hotpink_glasssunset · reddit · 2026-11-01
A developer's post-mortem: crashing agents are easy to fix; the ones that quietly ruin your week return confident, well-formatted output that is subtly wrong — a date off by one, a total missing tax, a summary dropping the one caveat that matters. These pass every smoke test. His mitigations: assert on the output itself rather than just checking it ran, verify all numeric results with a second cheap independent pass, and require factual claims to cite the source row. He argues the almost-right-but-not-quite failure is the real tax on agents, and no model upgrade has eliminated it.
More from coding & agent
- DeepSeek design doc: a person can only follow 3-5 agent sessions at once, attention is the real bottleneck — heyshrutimishra · 2026-11-02
- Software fundamentals are key to getting good results from AI coding agents — bendee983 · 2026-11-02
- Open-source LLM regression gate digline hits 1.0: quantifying judge noise before calling a regression — alexpran · 2026-11-02
- PR adds patient retry logic for model capacity errors in OpenAI Codex — bytebot · 2026-11-02
- Perplexity open-sources WANDR: 500 research tasks, 170,495 evidence points, URL-rechecking grader — thisguyknowsai · 2026-11-02
- 4 AIs, 96 questions, 2,671 citations: 74% are blog posts, 99% on companies' own sites — tibo_maker · 2026-11-02