EvoUndo: 197 of 600 agent self-edits fail recovery test

Across 600 self-evolution tasks, an agent modifying its own harness produced 197 capability-improving mutations that could not be safely undone. Conventional repair strategies recovered zero of them.

The finding sits in a preprint called "EvoUndo: Recoverability-Constrained Self-Evolution for LLM Agent Harnesses", posted to arXiv on August 28, 2026 by Tanmay Sah, Dolly Sah, Harshul Jain and Tanya Sah, listed as independent researchers. Their setup is agents that "modify their own prompts, tools, middleware, resources, and execution harnesses at runtime." Those mutations sometimes make the agent better. They can also leave persistent effects "that cannot be safely reversed in states different from the one in which it was created."

The 197 failures came from 600 one-shot tasks. Under the original recovery representation, iterative prompting-based repair got 0/197 back. A deterministic oracle, given the original recovery language L₀, reached 48/197. Expanding the recovery calculus pushed the oracle to 191/197.

The paper then splits the remaining problem in two. Exact state-address grounding lifted recovery from 0/48 to 38/48, or 79.2%, when the language was already sufficient. Extending the language itself reached 142/143 (99.3%) in the oracle-defined S₁ stratum. On the primary gpt-oss-120b backbone, adding exact-address diagnostics on top of the richer language actually reduced recovery to 133/143 (93.0%); a Qwen3.8-27B replication preserved the grounding and expressivity effects but not this negative interaction, "indicating that the latter is model-dependent."

The conclusion is blunt: "reliable agent self-evolution requires co-designing verification, state grounding, witness semantics, and recovery-language expressivity rather than relying on iterative prompting alone."

One paper, one benchmark, one primary backbone. The abstract publishes headline counts but no per-task breakdown of which capability-improving mutations turned out to be the unfixable ones.