Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-31 · high
Standard rollback prompts recovered zero of the 197 that stuck, but grounding the exact system state fixed 38 of 48.
Why now: The paper posted to arxiv in August 2026, putting a number on rollback risk for agents that edit themselves while running.
EvoUndo tested whether an agent can reverse changes it makes to its own prompts, tools, middleware and harness while it's running. Researchers logged 600 of these self-made edits. 197 improved the agent's capability enough that they couldn't be reversed under the system's original recovery instructions.
That's the risk for anyone building a harness that lets an agent modify its own configuration on the assumption a bad edit can be rolled back. Under that original recovery language, the EvoUndo paper recovered zero of the 197.
The researchers ran a 2x2 test to isolate why. Grounding the recovery process in the exact system-state address it needed to target lifted success from 0 of 48 to 38 of 48. That held in the slice of cases where the original recovery instructions were otherwise sufficient. Extending the recovery calculus beyond fixing the addressing reached 142 of 143 in a separate, oracle-defined stratum.
Each link below shares sources, entities, or timing with this story.
Same source domain / Shared topic / Tension / Downstream implication
Reported by the same outlet (arxiv.org); overlapping topics (agent, cannot); pushes against this story (but).
Same source domain / Shared topic / Downstream implication
Reported by the same outlet (arxiv.org); overlapping topics (agent, capability, repair); traces where this leads (downstream).
Same source domain / Shared topic / Tension
Reported by the same outlet (arxiv.org); overlapping topics (agent, harness, repair); pushes against this story (but).
Reported by the same outlet (arxiv.org); overlapping topics (agent, edit); pushes against this story (against).
Reported by the same outlet (arxiv.org); overlapping topics (agent, capability); pushes against this story (but).
Reported by the same outlet (arxiv.org); overlapping topics (agent, harness); pushes against this story (against).
Reported by the same outlet (arxiv.org); overlapping topics (agent, cannot); pushes against this story (against).
Reported by the same outlet (arxiv.org); overlapping topics (capability, exact); pushes against this story (against).