Agent kept searching but never opened the source: four runs expose a hidden failure mode
memokris · reddit · 2026-09-23
The author diagnosed a subtle agent failure: a task required "search, then use the returned ID to read the original text." Both tools were available and the prompt spelled out the sequence, yet in four recorded runs the second call never happened.
- Broad searches ate the call budget, sometimes on near-duplicate queries
- Once the budget was gone, finding something worth opening didn't matter
- Fix: restrict broad searches in later loop iterations, reserving room for dependent calls
Key lesson: don't judge by the final answer alone — the trace must show an actual source lookup, not just another plausible search result. Four runs aren't enough to generalize across models, but enough to change how they validate agents.
More from coding & agent
- Claude Opus system prompts keep growing: Opus 5.5 hits 4,108 words, mostly refusal guidance — rajistics · 2026-09-23
- WorkSwarm pitches real agent teamwork: persistent sessions, model routing and RSI — HeyToha · 2026-09-23
- Open-Source WorkSwarm Builds AI Agent Teams With Persistent Sessions and Recursive Self-Improvement — HeyToha · 2026-09-23
- Yacine asks: is anyone building an RL environment for code smells? — yacineMTB · 2026-09-23
- Qwen3.8-Flash Free in Qoder: Zero Credits for Individual Users, 14-Day Pro Trial With 300 Credits — mhdfaran · 2026-09-23
- Qoder walkthrough: start with one outcome, skip juggling five tools — mhdfaran · 2026-09-23