Your agent isn't dumb, it's just multiplying: 95% per-step reliability drops to 36% over 20 steps
Infinite_Web1576 · reddit · 2026-11-10
A Reddit post does the boring math on agent failures: at 95% per-step reliability, a 5-step chain succeeds 77% of the time, 10 steps 60%, and 20 steps just 36% — better prompts can't fix probability. Practical mitigations: keep tasks small ("pull 4 fields from this invoice", not "handle my finances"); validate output after every step, even simple format checks, to catch errors before they snowball; and require a human click for anything irreversible like emails, payments, or deletions. The mindset shift: stop maximizing autonomy and instead list which steps you're okay with the agent silently messing up — that list is shorter than expected.
More from coding & agent
- Agent swarm experiment: prompt name order and model choice decide who leads — bibryam · 2026-11-10
- Botzskool MCP server lets your AI agent loose on a virtual playground — EnvironmentalRice348 · 2026-11-10
- Your agent is slow — tracing shows the model usually isn't the bottleneck — aureliatech · 2026-11-10
- OpenCode v2.2.7 ships GitHub Copilot Auto routing, tab shortcuts and SDK breaking changes — github-actions[bot] · 2026-11-10
- Skills vs linked docs: where should AI agents' knowledge live? — IllustriousFault6218 · 2026-11-10
- Fable 5.2 models a Bugatti Tourbillon in Blender from a single prompt in 2.5 hours — socialwithaayan · 2026-11-10