Paper: https://t.co/wXZ6KF2lKm
Key pages
p. 2 the whole loop on one diagram
p. 5 how the yes or no checklist is built before each task
p. 6 the three lists (broke, never solved, mastered)
p. 7 the gate that decides if an edit ships
p. 12-13 results, incl. the 54.5% token cut
p. 26 the actual prompts for the checklist writer and the self-reviewer
Google just published the best paper on self-improving agents I've read this year.
Their harness rewrites its own rules after every run, never forgets what it learned, and uses 54.5 percent fewer tokens.
The model stays the same. Only the rules file around it changes.
Normally self editing agents get worse over time, because a fix for today's failure breaks what worked yesterday + the file keeps growing. GUARD solves this by tracking what each edit broke.
How it works.
1. Before the task starts, the agent turns it into a short checklist of things a good result must have. Each check either passes or fails.
2. After the run, the agent goes through that checklist and marks every check as passed or failed. So it learns what went wrong on it's own.
3. It runs the same tasks with the old rules and the new rules and sorts every check into three groups:
> Broke after the last edit. It passed with the old rules and fails now. These get fixed first.
> Never solved. It failed both times.
> Already working. These rules get locked so later edits can't touch them.
4. A rule that broke something isn't deleted. It gets an "only if" condition so it fires only where it belongs.
5. It writes several new versions of the rules in parallel and keeps the best one.
6. That version ships only if it helps more tasks than it hurts. Rejected edits go into a log so they never come back.
Results. Zero forgetting across four shifting task streams, while other methods fell below the starting agent. The rules file came out 2 to 4 times shorter. Tokens per task dropped from 131,032 to 59,651.