23 days of an agent picking its own tasks: 41 runs, 19 shipped, and why the 54% failure rate is the feature
PretendLime6041 · reddit · 2026-09-03
The author ran an autonomous agent loop for 23 days: a GitHub Actions workflow fires at 6:00 AM daily, reads search/usage data plus a log of every past initiative (outcomes, costs, failures), generates and ranks candidate tasks, commits to exactly one, then plans, implements, tests and verifies.
The core mechanism: nothing ships unless it passes 81 automated checks. Failed runs die quietly and get logged, feeding the next morning's task ranking — the only place learning happens in the system.
Tally from Aug 11 to Sep 2: 41 runs, 19 reached production, 22 died (54% failure rate). The author argues this is the feature: tuning for high success would mean loosening checks and returning to manually reviewing every diff. Of the 19 shipped, 7 changed the loop's own machinery, 12 were content pages; every word was agent-written, zero edits.
Two arguments: autonomy isn't a property of the model but a threshold you set with checks, not prompting. And the unsolved problem: task ranking is the only step with no failing test — a bad choice that clears all 81 checks ships exactly like a good one. The author asks how to evaluate the decision step, not the execution.
More from AGI Musings
- An estimated 30-40% of TikTok videos about the Lindsay Clancy trial are AI fakes — juliey4 · 2026-09-03
- Once consciousness's functions are explained, is the hard question solved? — ctjlewis · 2026-09-03
- If interpretability works, RL training could self-correct detected failure modes — mayfer · 2026-09-03
- Six Years Ago This Station Bookshop Had No AI Books — Now They're Everywhere — Merzmensch · 2026-09-03
- 1,284 agent jobs overnight for $4.98: Grok Bot beta receipts show the one-employee company — anirbanbandyo · 2026-09-03
- Chollet: all AI converges to symbolic learning as 8-year paper finds implicit symbolic structure in LLMs — burny_tech · 2026-09-03