Microsoft Paper Shows LLM Agent Success Rates Decay Geometrically Over Long Workflows

A Microsoft paper systematically measures how LLM agent performance decays over long workflows: with per-step error probabilities compounding geometrically, success rates across 9 models drop to 0-33% after 16 steps.

2026-09-08 ~ 2026-09-08 · 2 related posts

• Microsoft paper: agents near-perfect at 2 steps collapse to 0-33% by step 16 — rohanpaul_ai · 2026-09-08
• arXiv: agent success follows a geometric decay law, collapsing within 16 steps — rohanpaul_ai · 2026-09-08