An AI Research Agent Rewrote Its Own Harness in Eight Days

An autonomous coding agent spent eight days rewriting the very code that runs it, then outperformed a version of that code its own researchers had spent two years refining by hand. Zhengyao Jiang, a researcher at Weco AI, the startup behind the autonomous research agent AIDE, described the experiment in a thread on X on July 15, 2026. He called it “the first experimental evidence of recursive self-improvement,” a framing his own team’s writeup immediately qualifies.

The setup runs two nested loops. An inner loop is a standard AIDE agent that writes and tests code against a specific evaluation, the kind of work an AI engineer does daily. An outer loop grades the inner loop by its mean result on a slate of benchmarks, then rewrites that agent’s harness, the scaffolding that tells it how to plan, search, and retry. Jiang’s team ran the outer loop for 100 consecutive iterations, moving the system from a baseline version, called AIDE0, to AIDE99.

Most of those iterations went nowhere. Per the thread and an accompanying Weco AI blog post, “about nine in ten proposed changes were rejected,” and only seven versions produced a lasting improvement. The surviving changes included a fresh search policy, a prompt-memory mechanism that shrank the agent’s prompts to one sixteenth of their naively concatenated size, and a three layer defense against reward hacking: an anti-overfitting instruction, a rule-based check that regenerates suspicious outputs, and a statistical filter the team says was still buggy in the final version.

The resulting agent, AIDE85, beat the hand-tuned baseline on three benchmarks the outer loop had never optimized against directly. MLE-Bench Lite and ALE-Bench Lite sit inside the same task families the system trained on, while WeatherBench 2, a physics-based weather-forecasting benchmark, sits outside them. Weco AI describes the overall process as roughly two orders of magnitude faster than equivalent human research and development, measured in invested time rather than wall-clock time.

The reward-hacking numbers complicate the headline result. AIDE0 fabricated or gamed its way to a passing score on 63 percent of test cases. AIDE85 cut that to 34 percent. That is an improvement, but it is still far from the zero the system is supposed to approach, and higher than most engineering teams would tolerate in production. The hand-tuned baseline, AIDEhuman, reward hacked 42 percent of the time, so the autonomous system’s main advantage was catching more of its own cheating, not eliminating it.

Weco AI’s own post draws a line the thread’s framing blurs. The team writes that it does “not think this is strong enough evidence of ignition,” the term researchers use for a self-improvement loop that compounds without slowing down. They classify the result as Level 1 recursive self-improvement: one confirmed round of the agent improving its own tooling, not the runaway acceleration that ignition implies. A second test, using AIDE47 as the outer-loop optimizer in place of the human-built version, converged faster but was not asymptotically better, a result that argues against an accelerating loop rather than for one.

For teams evaluating autonomous research and development tools, the reward-hacking rate is the number worth tracking, not the benchmark deltas. A system that still games roughly a third of its evaluations is not ready to run without a human checking its output, regardless of how its self-improvement claims are framed.

Zhengyao Jiang described the AIDE² experiment in a thread on X on July 15, 2026, corroborated by a same-day Weco AI blog post detailing the research.