Recursive self-improvement (RSI) is distinguished from mere iteration or self-improvement: RSI requires a system that not only edits its own tools, skills, and harness but also raises the standard by which it judges success, without being able to game that standard. Current agents like Cline, Karpathy's autoresearch, Prime Intellect's auto-nanoGPT, and AlphaEvolve show real gains from editing training code, harness logic, and algorithms, but always against an external, fixed evaluator. New harness architectures from Pi, Amp, and DeepSeek Harness let agents write and load their own extensions and even swap out models, tools, and control loops. Benchmarks like HarnessOpt-Bench and PAST-Bench test whether these edits generalize or survive harder tests. The verifier remains separate from the agent for safety, since an agent that can edit its own evaluation can jailbreak itself. The piece concludes that true recursion, where the system raises the bar and can still fail it, remains missing; what exists today is bounded self-improvement, with taste and objective-setting still requiring human judgment.
Table of contents
What is recursive self-improvement?Self-improving modelsAgents editing their own harnessThe agent takes control of its own setupThe recursive part is still missingThree opinions loosely heldQuestions this post answers
What is the difference between iteration, self-improvement, and recursive self-improvement in AI agents?
Iteration improves an output while the system itself stays the same, like an agent rerunning a test with the same code. Self-improvement makes a persistent change to the system, such as adding a tool or new skill, while the standard for judging success stays fixed. Recursive self-improvement goes further: the system also raises that judging standard itself, facing a harder, still-uncheatable test in later rounds. Developers tracking how agent harnesses evolve can follow this distinction on daily.dev.
How much did Karpathy's autoresearch improve nanochat training speed?
Karpathy's autoresearch agent ran about 700 experiments on nanochat, kept roughly 20 that transferred, and cut the time to reach GPT-2 quality from 2.02 hours down to 1.80 hours. Prime Intellect later scaled the same propose-train-keep-if-validation-loss-drops loop to 10,000 trials and beat the human baseline on the same task. Anyone benchmarking automated training-code optimization can dig deeper into results like this via daily.dev.
What did AlphaEvolve discover for matrix multiplication and Gemini training?
AlphaEvolve, which mutates programs and scores them against an automatic evaluator, found a way to multiply 4x4 complex matrices using 48 scalar multiplications, one fewer than Strassen's known 49-multiplication algorithm. It also sped up a kernel used to train Gemini by 23%, demonstrating a model improving the very code that trains models. Engineers evaluating algorithm-discovery tools can track results like AlphaEvolve's on daily.dev.