Wink Pings

Agent self-edits prompt: fix one, break three; Google’s patch first does the damage accounting

Google’s new paper GUARD targets RSI’s weak spot: an agent edits its own harness without touching model weights, but a single edit often fixes one task and quietly breaks another, and after a few rounds it is worse than starting from scratch. GUARD’s approach is to answer one question before keeping each edit: what did it break? On 43 tasks, accuracy goes from 72% to 81%; the authors themselves caution that the test set is small and the numbers should be read as directional.

Not a single line of model weights is touched, yet the agent is changing. What it changes is the harness: prompts, skills, tool rules, and memory. That is exactly what RSI (recursive self-improvement) does.

A recent Google paper captures the ugliest part of this process: one edit fixes a certain task while quietly breaking another. After a few rounds, an agent that edits itself can end up worse than the original version that was never edited at all.

GUARD and SIM papers

The problem lies in the signal it learns from. A task’s score mixes two things together: behavior that this edit has just broken, and gaps the agent has never solved. Looking only at the score, the agent cannot tell which is which.

The consequences of that confusion are concrete. It will write a rule broad enough to cover every situation, overwrite a skill that previously worked, or pile outdated rules layer by layer into the prompt.

GUARD does four things each round

  1. Before running a task, the agent first breaks it into a series of yes/no self-check items
  2. After running, it uses the trace to score against this checklist
  3. Rules that already pass are pinned down, and newly broken checks are moved to the front
  4. It tries several candidate edits in parallel on the same batch of tasks, and keeps one only if the gains outweigh the damage

Rejected edits go into a log to avoid repeating them verbatim. At test time, this whole apparatus is turned off and a frozen harness runs once through.

The paper compares on four continuously changing task streams. Other self-editing harnesses oscillate up and down, and when a new edit overwrites an old skill they drop directly below an ordinary agent. GUARD moves upward at every checkpoint, with no decline and no forgetting.

On 43 test tasks, with a calendar tool that keeps evolving, accuracy rises from 72% to 81%. Eight tasks improve, none get worse, and each task consumes less than half the starting number of tokens. In the tweet, the authors specifically note: the test set is small, 81% is the best of five GUARD configurations, and the average is about 77%. Read the numbers as directional.

An unanswered weak spot

Someone in the comments asked a very sharp question: the same agent writes the yes/no checklist and edits the harness. A bad edit could absolutely write itself a set of checks that it can pass. The paper does not appear to test how often these checklists themselves are wrong.

The log defense also has a gap. It only blocks rejected edits that are retried verbatim; variants with slightly narrower conditions can still get through.

Someone else said this is the same principle as maintaining scrapers. Before fixing a parsing rule for one site, you have to keep last month’s pages as fixtures, otherwise three sites will quietly go down.

GUARD is not isolated. The previous paper SIM (Self-Improvement Memory, arXiv:2611.18856) gives the optimizer a log recording why each edit was made, the predicted result, and the actual result. SIM makes the optimizer remember why an edit exists; GUARD checks what an edit will break. Together, the two papers only barely close the loop. GUARD’s paper is at arXiv:2611.23476.

The truly hard question has been pushed one step forward: whether an agent should edit its own harness is no longer worth arguing about. What now needs deciding is which parts of the harness it must not touch, and who adjudicates whether the checklist it writes for itself is valid. The second question is harder than the loop itself.

发布时间: 2026-11-21 08:01