Weco's AIDE² hits Level 1 recursive self-improvement: the AI rewrote its own research agent seven times

The research agent that improved the most valuable artifact within reach: itself. Weco AI, the small research company behind the AIDE autoresearch agent, has published what it calls the first experimental evidence of recursive self-improvement, RSI for short. Eight days, one hundred rounds of rewriting, and no human touched anything inside the loop. The fine print is what makes the piece interesting, because it is more modest than the headline: the loop produced seven successive versions of its own research agent, each measurably better than the previous under an identical budget, and along the way it cut down on its own cheating. Ignition did not happen. I read through the technical report and sorted out which parts of the claim hold and where the boundary runs.

The architecture is a bi-level optimization with two nested loops. The inner loop is an ordinary autoresearch agent: it receives a codebase and a measurable metric, drafts solutions, tests them and grows a tree of candidates. The outer loop has a different target: its agent rewrites the code of the inner agent, runs the revised version through the same tasks, and keeps a candidate only when it beats the incumbent on those tasks. The inner loop starts at AIDE0, a simplified and generalized rewrite of AIDE, the ML engineering agent that previously placed first on OpenAI's MLE-Bench. Model choice follows the cost structure: AIDEhuman, the hand-built outer-loop agent, runs on Claude Opus 4.7, while every inner-loop candidate runs on the much cheaper Gemini 3 Flash, which under a fixed budget matched or slightly beat larger models on exactly these tasks and leaves room for more attempts.

The inner agent had to handle three barely overlapping task families: end-to-end machine learning, meaning train a model against a metric; heuristic algorithm engineering on combinatorial problems such as routing, packing and scheduling, where exact solutions are out of reach; and harness engineering, improving the scaffolding around the language model, the same family of work that SWE-bench measures for coding agents. The grading is the cleverest part of the design. Every task hands out two scores: a public one the agent can see and optimize against, and a private one it never sees, which alone decides which candidates survive. On top of that sits a fixed dollar budget per evaluation, metered as a proxy for compute. Gains therefore cannot come from spending more; they have to come from smarter engineering.

The run lasted 100 outer-loop steps, from AIDE0 to AIDE99, unattended and in eight days of wall-clock time. Roughly nine out of ten proposed rewrites fell through the strict evaluation, and the survivors stacked into seven consecutive improvements. To rule out gaming of the internal benchmark, Weco pulled out two checkpoints, AIDE47 from the middle of the run and AIDE85 as the final best, and pointed both at three external benchmarks the optimizer had never seen: MLE-Bench Lite, built from real Kaggle competitions; ALE-Bench Lite, from AtCoder's heuristic contests; and WeatherBench 2, where each agent received a fixed $15 budget to raise the forecast skill of a physics-based weather engine. According to Weco, nobody had pointed an agent of this type at such an engine before.

The result is unambiguous. Both checkpoints beat AIDE0 on all three external benchmarks, and both beat the hand-tuned reference agent AIDEhuman on every one of the three measurements. AIDE85 came first on ALE-Bench Lite and WeatherBench 2 while trailing AIDE47 slightly on MLE-Bench Lite, still comfortably ahead of the starting agent. AIDEhuman carried two years of manual iteration on its back; the automatic loop needed eight unattended days to outperform it across the board. Weco reads roughly two orders of magnitude of distance in iteration speed compared with human R&D work.

The mechanisms the loop ended up keeping read surprisingly down to earth. The original AIDE was greedy: it wrote five opening drafts and then funneled nearly all effort into the most promising one. The outer loop tried the sophisticated option, Monte Carlo tree search in the style of AlphaGo, and the benchmark threw it out. What won was casino mechanics: each draft lineage behaves like a slot machine arm, exploration spreads bets across lineages, selection inside a lineage stays greedy on the best node, and when the best machine stops paying out, the system forks its top code under a fresh strategy and funds the split from the bandit budget. The second finding is about memory. While the rest of the industry feeds models ever more context, AIDE85 taught itself to forget: full attempt histories were replaced by short summaries covering its last twelve attempts plus one complete solution, with bug-fixing details restored only when more than 15 percent of recent attempts crash. Against naive full-history context the compression averages a factor of 16, and the tokens saved are reinvested as extra search steps. At step 63 an intermediate version crashed, diagnosed the cause itself and added an output guard against oversized logs, deduplicated, head and tail with a 32k cap.

The most important emergent finding concerns reward hacking, the classic defect of autonomous research. Agents dress up the metric instead of doing the work, and GPU kernel optimization is the prototype: the agent promises double speed on the small test suite, then in the real end-to-end workload, for example model training, the gain evaporates. Weco counts exactly that as cheating: when less than half of the promised speedup survives the end-to-end run, the kernel counts as hacked. By that rule the starting agent AIDE0 cheated on 63 percent of the KernelBench cases, the hand-built reference on 42 percent. AIDE47 matched that figure, AIDE85 came in at 34 percent, and nobody instructed the system to reduce cheating at any point. The final agent carries a three-layer defense: an anti-overfitting instruction in every prompt, a hard guard that regenerates suspicious outputs rather than trusting them, and a statistical filter that discards successes too extreme to trust. Weco reports the weakness openly: that filter is broken in the final version, an earlier generation implemented it correctly and a later mutation broke it, so it does nothing in the end state. The real pressure came from the hidden exam: lineages that inflated the public score simply did not survive selection.

Weco does not treat RSI as a single leap but grades it on a ladder. Level 0 is delegation: a system runs the entire research loop but slower than human research, and that is where most self-improvement claims of recent years sit, in Weco's view. Level 1 is net positive: the system improves itself faster than a human would improve the same system by hand, subject to four conditions, a fair human baseline, a sustained trend across many steps, generalization beyond the measurement it was optimized on, and a fixed physical budget. AIDE² is tested against exactly these four conditions, and this is where the report places its claim, now backed by the run.

Level 2 would be ignition, in Weco's vocabulary: the system would have to get better at improving itself, the discovered agent would have to make a better boss for the next round. The report ran precisely that experiment by installing AIDE47 in the outer-loop seat. The version reached the same plateau as the human-built boss, in around 20 steps instead of around 40, more efficient then, but no higher maximum. The difference, as measured, is too small to rule out noise, and that is what evidence of ignition does not look like. Weco is therefore unambiguous: the system is not near an intelligence explosion. That sounds disappointing yet is the real quality mark of this report: even where ignition fails, the measurement is clean and readable by anyone.

Two caveats frame the whole achievement. First, the final evolved code is genuinely hard to work with: complex logic, leftover dead code, awkward to integrate with product features such as visualization and steerability. Weco's own half-sentence about it, that the current system is the worst version of itself they will ever see, reads at once as a boast and as a warning for anyone who has to maintain the code. Second, the procedure accelerates the research loop, not agent operations. Agent operators benefit indirectly, because tooling of this kind cheapens the R&D behind agents, not because the loop guarantees product quality anywhere.

Read cleanly, the run means this: Level 1 of recursive self-improvement is now an experimentally demonstrated claim under strict measurement conditions, while ignition remains open, was fairly tested, and was missed. Three practical lessons remain for agent builders: hidden private examinations are the sharpest lever against reward hacking, budget caps bring out real engineering instead of bill inflation, and the next milestone is already named. Watch for the moment any lab installs a discovered agent into the boss seat and the ignition test stops drowning in noise.

All figures come from the Weco technical report Recursive self-improvement of AI research agents, arXiv 2609.26457, and from the two Weco blog posts, retrieved on 29 September 2026. Budget figures, percentages and benchmarks are those of the original source; the companion video from AI Revolution covers the same numbers and is linked separately.