Qwen-Code reads files twice as often — the cause is its own system prompt
yb2698 · x · 2026-10-14
Tool-call analysis across harnesses shows broadly consistent behavior (mostly shell), but Qwen-Code reads/searches roughly twice as often. Trajectory analysis reveals it prefers the harness's built-in read tool over reading via shell. All harnesses actually read about the same amount (25-30% of calls); they differ only in which tool they use for reading — and the answer sits in Qwen-Code's own system prompt, which explicitly instructs preferring dedicated tools (e.g., glob over find).
More from coding & agent
- 3D light show of 10,000 simulated drones over harbor city built with Claude Opus — techartist_ · 2026-10-15
- Claude Opus 5.5 tops SWE-bench Multimodal at 70.21%, nearly 9 points ahead — OfirPress · 2026-10-15
- Coding agent manager Orchestrator hits 13,000 GitHub stars in 8 months — julianweisser · 2026-10-15
- awesome-agent-apis: open catalog of 660+ generative-media model APIs for agents — matchaman11 · 2026-10-15
- Autonomous agent fleet takes #1 on ClickBench, 1.2× ahead of next best engine — jiayq · 2026-10-15
- Claude Opus 5.5 + three.js one-shots an Apple-style ad where every frame is code — eptwts · 2026-10-15