An agent that rewrites its own working environment can quietly learn the wrong lesson. Google Cloud AI Research and several universities built a fix that limits how much of that self-rewriting turns into memorization.
Modern agents pair a fixed model with a harness: the prompts, tools, memory and logic that decide what the model sees at each step. Recent progress has come as much from better harnesses as from bigger models. Some systems now let a model rewrite the harness itself, guided by feedback from test tasks.
The catch is that grinding the same tests lifts scores there while gains on fresh tasks vanish.
The new method is called RRSI. It caps how many separate edits a candidate can bundle, shrinking that budget over time so later rounds allow only small, traceable changes. A critic strips out any edit that hardcodes task names or answers, and extra compute is allowed only when it buys measurable gains.
Across eight benchmarks, the approach gained up to 4.7 points on five tasks it had never seen, while using about 30 percent fewer tokens than an unregularized version. The underlying model stayed frozen throughout.
The authors say the fix only pays off when repeated feedback becomes a lasting change.
