The compounding error problem: 95% per-step accuracy means 36% over 20 steps
KelvenSevero0 · reddit · 2026-10-28
A developer on agent reliability math: at 95% per-step accuracy, end-to-end success drops to 77% over 5 steps, 60% over 10, 36% over 20—and failures are rarely clean crashes, just confidently wrong turns noticed three steps later.
His landing point: code owns the control flow, the model owns judgment calls—LLM handles fuzzy parts (classify, extract, decide, draft), every output is validated by plain code before the next step, and the model picks its own step only when the path can't be known in advance. Rule of thumb: if you can draw the flowchart on a whiteboard, it shouldn't be fully agentic.
Full autonomy works when the path is unknown upfront, a human reviews the result anyway, and wrong turns are cheap—coding agents tick all three (diff review, tests, git revert). He asks: who runs fully autonomous agents in prod with no human in the loop, and how do you measure per-step reliability?
More from coding & agent
- Matt Pocock mulls banning useState outright in his React apps to steer AI coding agents — mattpocockuk · 2026-10-28
- Developer: tickets, code, tests and PRs should all be generated by agents — intellectronica · 2026-10-28
- Simon Willison: October 2026 is the vibe-coding-games inflection point, Opus 5.5 knows fun — Simon Willison · 2026-10-28
- Avalon-M1 trains LLM agent organisations via collective-reward RL, lifting avg score 33.5% to 45.7% — RobertTLange · 2026-10-28
- Creator chains Halo and Runway via MCP connectors in GPT for a music video workflow — aziz4ai · 2026-10-28
- Dev uses Qwen in an opencode harness to root an LG TV and strip out spyware — Switchboy1 · 2026-10-28