Optimizing Coding Agents the Wrong Way: A Fail-Closed Chain Runs 26 Days Autonomously
nodo48 · reddit · 2026-08-28
The author argues the industry optimizes coding agents for the wrong goal: everyone chases speed (more context, tools, autonomy), but almost any agent looks reliable in a 20-minute run. The real question after thousands of decisions: who validates whether the agent had the right to advance to the next governance phase?
So he built the opposite: a fail-closed chain designed to transition autonomously through any program. It currently runs an 8-milestone epistemic program on a real repo:
- Running 26 days 14 hours, hundreds of commits, closing stages with primary evidence, discarding strategies and reopening "finished" work
- Key result: after nearly a month, the agent's errors have no authority to declare success or skip pipeline phases
- Evasion paths found by adversarial testing were converted into hostile cases it must pass to unlock the next step
The implementation stays private; he asks the community for comparable runs of autonomous meta-governance chains over weeks without daily human direction.
More from coding & agent
- Visualizing Invisible AI-Driven Workflows — DavidWells · 2026-08-28
- ChatGPT + ThreeJS Mock API Boosts Graphics Dev Efficiency — collinalexbell · 2026-08-28
- OpenResearch Launches AutoResearch: Automating Paper Replication with Agent Swarms — simonguozirui · 2026-08-28
- Merge Gateway update: per-project guardrails, Parallel web search, audio/video multimodal — shensi · 2026-08-28
- End-to-End Legal Agent Guide: Zero Hallucination & Human Handoff — MaryamMiradi · 2026-08-28
- OpenInstinct: Self-hostable iMessage AI assistant with browser control — arthurcolle · 2026-08-28