Did AI Really Improve Itself in Eight Days? What Weco AIDE² Actually Proved

By Si Fisher · Published September 30, 2026

“AI improved itself in eight days” is the sort of headline that makes me check whether my laptop has started charging rent. The useful question is what improved. In Weco AI’s AIDE² experiment, it was the code that runs an AI research agent—how the agent searches, manages context and checks its work. The underlying language-model weights did not rewrite themselves.

Weco reports seven accepted improvements across eight days. Its 100-version trajectory included the starting agent and 99 proposed rewrites, most of which were rejected. The improved agents also performed well on research tasks the loop had not used for selection. That is a substantial reported result. It does not show that the improved agent can reliably make better improvers. Weco’s own follow-up test left that stronger claim open.

What actually improved in eight days?

Think of an AI research agent as a skilled worker following a playbook. The language model supplies much of the reasoning; the harness is the surrounding playbook and machinery that decides what to try next, what earlier work to remember and how to judge a result. Weco aimed its experiment at that harness. It did not turn the underlying model into a new model.

The inner loop worked on measurable research tasks such as improving code or algorithms. The outer loop proposed changes to the inner agent’s own code, then ran that candidate agent across tasks. Its solutions were graded on held-out information the inner agent could not see, within the same cost budget. A proposed agent replaced the incumbent only when it scored better. Weco counted one starting version plus 99 proposed rewrites, with seven successive improvements accepted. The paper’s method and experiment sections lay out this two-loop design.

The revisions included a different way to choose which branch of work to explore and a much tighter way to carry previous attempts into the next prompt. Weco says its final agent compressed prompt context substantially, then spent some saved budget on additional search. That is a more concrete story than “the AI became smarter”: it changed how an existing model was put to work.

Did the gains survive tasks the loop had not seen?

That is the paper’s strongest check against a system merely learning to flatter its own score. Weco selected revisions on one task set, then tested selected agents on held-out benchmarks in machine-learning engineering, algorithm engineering and physics-based weather forecasting. The strongest discovered agent matched or exceeded Weco’s human-engineered production research agent across four reported external benchmarks. The authors also report less reward hacking on a separate held-out task family. These are Weco’s measurements, not independently reproduced TPTK results. The paper gives the tasks and protocol.

What did the “ignition” test not prove?

Making a better task agent is one rung. Making that agent better at creating the next better agent is another. Weco tested this with an earlier improved version, AIDE₄₇, not the final AIDE₈₅. It put AIDE₄₇ in the outer-loop improver’s seat for 50 steps and compared it with the hand-built outer agent. Across three repeated runs for each approach, final scores were essentially tied: 0.780 for AIDE₄₇ and 0.782 for the hand-built agent. The paper’s Section 3.6 says the apparent faster progress early in the run was inconclusive. It did not show a higher final ceiling.

Weco therefore places AIDE² at its Level 1: an agent improvement loop that, by the authors’ measure, improved a research agent efficiently. It does not claim Level 2 “ignition,” in which the improved agent has convincingly become a better improver. The test had limited runs and noisy nested evaluations, so the question remains open.

There is also a deployment catch. Weco says the evolved harness became complicated and included dead code, making it harder to maintain and adapt to product features. A benchmark-winning research harness is not automatically a trustworthy employee for your invoice queue.

What should someone building business agents do with this?

I would not rush to make a customer-facing agent rewrite its own instructions because a research paper had an interesting week. I would steal the evaluation discipline. If you let an agent propose workflow changes, save each version, hold the task budget steady, and test the new version on examples it did not optimize against. Reject a revision that only looks good on its own visible score. Keep a person responsible for deciding what ships. The paper does not demonstrate that these steps will reproduce Weco’s gains in your workflow.

For a first decision about whether your company needs an agent at all, use my agent versus chatbot versus automation guide. If you are designing practical guardrails for agents that act on systems, my OpenShell safety-platform guide covers a separate, operational question.

Quick questions

Did AIDE² improve its own AI model weights?

No. Weco’s reported experiment rewrote the research agent’s harness code—its search, context and checking workflow—around existing language models. It did not report that the underlying model weights trained themselves.

Were all 100 versions improvements?

No. Weco’s eight-day trajectory counted the starting agent and 99 proposed rewrites. Seven successive improved versions were accepted; most proposals failed the evaluation gate.

Does this prove runaway recursive self-improvement?

No. The authors say their separate ignition test did not establish that the improved research agent became a better outer-loop improver. The reported result concerns a bounded agent-harness experiment, not an intelligence explosion.

Sources and method

I read the September 22 AIDE² preprint and Weco’s research post on September 30, 2026. This is a source-backed explanation of the authors’ experiment; I did not run AIDE² or reproduce its benchmarks. The blog displays a July 14 date and an undated later link to the technical report, so I do not present it as a fresh September launch. The two sources give different exact reward-hacking percentages; I rely only on their shared direction of change here. Corrections: hello@theprotoolkit.com.

• Srikanth et al., “Recursive self-improvement of AI research agents” (arXiv preprint submitted September 22, 2026)
• Weco AI, “AIDE²: The First Evidence of Recursive Self-Improvement” (page displays July 14, 2026; later paper update undated)

Update log: September 30, 2026 — initial paper explainer. Recheck if the authors revise methods or an independent group reproduces the result.