Compound Mistake Amplification

Per-step accuracy multiplies across a multi-step agent trajectory, so long plans fail even when each individual step looks reliable.

In multi-step LLM agents, errors compound. If each step is 95% accurate, ten steps yield roughly 60% overall success; a hundred steps fall near 0.6%. Long tool-using trajectories therefore demand stronger models and tighter plan validation than single-shot generation.

Concrete edit loops show the same curve: any edit attempt may still succeed (~90% eventual recovery chance in SWE-agent GPT-4 Turbo runs), but after one failed edit recovery odds crash (~57%), and unresolved tasks are dominated by incorrect/over-specific implementations (~52%) plus cascading failed edits (~23% failed edit recovery). Successful recoveries usually follow ~2 consecutive failed attempts; unsuccessful ones linger ~4–6 failures before abandoning the edit. Successful runs finish earlier and cheaper; raising max budget rarely rescues the long failing tail — design for early localization and guardrailed edits, not endless retries. Transition patterns that matter for ACI design: reproduce via create→edit→python, then localize (find_file/search_dir→open→search_file→goto), then edit↔python loops; cascading edit×4 or scroll_down×4 is a smell.

This is one reason agents are higher-stakes than non-agentic use cases: tool access makes each wrong step more damaging, and multi-step runs cost more time and money. Design responses include shorter plans, stronger planners, plan-validate-execute separation, reflection and error correction, and human approval gates on irreversible actions.

---

Sources

• AI Engineering (Chip Huyen), ch. 6
• SWE-agent (Yang et al.), pp. 1–15 (§5.2)