EvoUndo: Recoverability-Constrained Self-Evolution for LLM Agent Harnesses
The paper finds capability gains can leave agent harness changes that cannot be safely undone.
EvoUndo tests self-modifications to prompts, tools, middleware, resources, and execution harnesses against recoverability across counterfactual states. In 600 unseen one-shot tasks, the authors found 197 capability-improving mutations that failed recoverability verification. Standard repair strategies recovered none of those failures under the original recovery representation, while an extended recovery calculus raised oracle recovery to 191 of 197. The authors argue reliable self-evolving agents need verification, state grounding, witness semantics, and recovery-language expressivity designed together, not just more prompting. ArXiv · AI/CL/LG's note
EvoUndo tests self-modifications to prompts, tools, middleware, resources, and execution harnesses against recoverability across counterfactual states. In 600 unseen one-shot tasks, the authors found 197 capability-improving mutations that failed recoverability verification. Standard repair strategies recovered none of those failures under the original recovery representation, while an extended recovery calculus raised oracle recovery to 191 of 197. The authors argue reliable self-evolving agents need verification, state grounding, witness semantics, and recovery-language expressivity designed together, not just more prompting. ArXiv · AI/CL/LG's note
score 5