AI agents can 'memorize' the test: a study curbs their self-optimization to prevent it

By DiarioBitcoin · October 4, 2026. A research paper —titled 'RRSI: Regularized Recursive Self-Improvement of Agent Harnesses' and tested with Google's Gemini models— puts numbers to a problem that agentic AI teams have long suspected: when an agent rewrites itself to improve its results, it can end up memorizing the exam instead of learning to solve new tasks. The target of the optimization is not the language model but the harness: the set of prompts, tools, memory and rules that decides what information the model receives and how it acts at each step.

The proposed method, dubbed RRSI, puts two brakes on that self-optimization loop. First, it limits how many changes the agent can bundle into a single proposal —a budget that shrinks over time, from broad changes at the start to small, traceable tweaks later on. Second, a 'critic' discards any modification that includes task names or solutions tied to a specific benchmark, and only allows the agent to spend more compute if that investment comes with a measurable performance gain. Across the eight coding, office-task and engineering-design benchmarks on which it was tested —with the model kept frozen to isolate the harness's effect— RRSI gained up to 14.1 points on the training tasks and up to 4.7 points on five benchmarks it had never seen, while consuming roughly 30% fewer tokens than unregularized self-optimization. The biggest jump on unseen tasks came on JobBench; two of the rival methods, by contrast, performed below the original harness as soon as they were put in front of a new task, the classic sign of overfitting.

The most practical finding may be a different one: a harness optimized with Gemini 3.5 Flash raised the accuracy of a weaker model, Gemini 3.1 Flash Lite, from 11.2 to 14.6 points on a coding test without touching its parameters. If that result holds up at greater scale, the harness stops being a fine-tuning of one specific model and becomes a transferable asset: an engineering layer that any team on a more modest budget could inherit instead of training from scratch.

In general, the underlying point matters more than the paper itself: much of the recent progress in agents comes not from new language models but from improving the harness around them, the 'how it is used' over the 'how smart it is'. It is the same logic that has for months explained the differences we see between labs in our own quality indexes: the model matters, but the scaffolding around it —memory, tools, error-recovery rules— accounts for much of the distance between an agent that works in production and one that only shines in the demo.

Our reading is that this work documents, with method, something we had already been flagging from the benchmark side: when the metric becomes the target, it stops being a good metric. What is called 'harness memorization' here is the agentic version of the saturation we already see in cybersecurity tests or in Elo scores optimized for voters: a system that learns to look good instead of being good. The difference is that RRSI offers more than a warning: a concrete way to audit the self-improvement process, ruling out shortcuts and requiring every unit of compute spent to be justified by a real gain outside the training set.

The limitation the authors themselves acknowledge matters: the study works with frozen-weight models, not with systems where model and harness evolve at the same time, which is where the frontier of agentic self-optimization is heading. Even so, the direction is the right one for our underlying thesis: if agentic AI is going to take on increasingly autonomous tasks on the way to the abundance we champion in the long run, it first needs mechanisms that distinguish real progress from the appearance of progress. An agent that only learns to pass the exam does not free up human time, it consumes it in supervision. RRSI's code is available on GitHub, which at least means that scrutiny does not depend solely on the word of whoever built it.