Self-evolving AI agents making irreversible changes
AccomplishedLeg1508 · reddit · 2026-09-02
As agents become more autonomous, they modify their own prompts, tools, middleware, routing, and execution harnesses. The core question: what happens when an agent makes a useful change but cannot safely undo it?
In work on EvoUndo, across 600 unseen tasks, 197 capability-improving mutations failed recoverability verification. Conventional repair recovered 0/197 under the original representation.
Experiments suggest bottlenecks are state grounding and recovery-language expressivity. The idea is simple: if an agent makes persistent changes, forward improvement isn't enough; the system must also verify safe recovery across different states.
Related event: EvoUndo: Restorability Constraints for Self-Evolving LLM Agents(5 posts)→
More from coding & agent
- Claude-written test harness let 31KB fake archives land on a real backup server — Grimmoner · 2026-10-18
- Matt Pocock's pre-commit hook caps agent-written files at 5,000 tokens — mattpocockuk · 2026-10-18
- Backups silently failed for 10 days: three monitoring blind spots lined up — Grimmoner · 2026-10-18
- Author Laurence Moroney: write your ugly first draft yourself, then let an LLM edit it — lmoroney · 2026-10-18
- Open-weight TYGER speech decision model makes 80ms turn-taking calls for full-duplex voice agents — solyarisoftware · 2026-10-18
- Devs: Opus 5.5 on medium beats Opus on max — eyishazyer · 2026-10-18