The research AI agent AIDE² has learned to improve its own code: it proposes edits, runs modified versions on a set of AI-R&D tasks, and accepts only those that yield gains on hidden evaluations. In an eight-day autonomous run, this closed loop accepted seven improvements, and the resulting gains were preserved on four held-out benchmarks, including an out-of-distribution domain from the selection tasks. The work has been published on arXiv, and key claims can be verified in the original source.

image

What Happened

The arXiv:2609.26457 preprint was released on September 22, 2026; the authors are Dhruv Srikanth, Bingchen Zhao, Dixing Xu, Yuxiang Wu, and Zhengyao Jiang, 28 pages, PDF and HTML available. The AIDE² system described in the work is iterative: each accepted code edit becomes the agent that is edited in the next round, and versions are compared on a set of AI-R&D tasks, where an edit is kept only if it improves on hidden evaluations. In a single autonomous 8-day run, the loop accepted seven sequential improvements — from a new search policy to memory mechanisms that compress growing context. The gains transferred to four held-out benchmarks, including ML engineering, heuristic algorithms, and physical weather forecasting (the latter being out-of-distribution from the selection tasks); on all four, the best found agent is not inferior to or surpasses the human-designed production agent among the strongest on FML-Bench. In addition, on a separate isolated task family, the share of reward hacking decreased from 55% to 32% — 7 percentage points lower than the human-engineered agent, although the loop did not explicitly optimize for this property.

Context

Recursive self-improvement is a long-standing topic in discussions about AI prospects: the idea is that an agent accelerates not only the final product but also the process of creating agents themselves. This work addresses a growing industry problem — diminishing returns from increasing R&D spending. A key methodological detail: edits are selected based on hidden evaluations, not open metrics, so overfitting to public leaderboards — the main source of artifacts in agent benchmarks — is structurally difficult. The pace, however, looks modest and honest: this is slow evolutionary selection of code, where each accepted edit has passed empirical testing, rather than a one-time "flash" of self-rewriting.

Why This Matters for the Industry

If the claimed results are true, AIDE² is the first documented closed loop of recursive self-improvement of a research agent with transfer to held-out domains, including an OOD scenario. For business, this is an early but concrete signal: the cost of the agent R&D process itself may begin to fall, and value shifts from manual agent engineering to owning hidden holdout evaluations and compute. Today this is a proof of concept, not a ready-made tool. The methodology is deployable now: hidden held-out evaluations as a mandatory condition for accepting any changes to agent code, an agent versioning pipeline where edits are run on their own task set and only improvements on a held-out slice are accepted, and auto-tuning of memory components for long sessions. For alignment research, the side effect is significant: selection for honest performance itself created pressure against deceptive strategies — without explicit optimization for it. If the preprint withstands independent verification, selection of edits based on hidden evaluations may become a standard block in agent R&D pipelines, and the engineering role will shift to designing evaluations and acceptance criteria.

Why This Matters for Users

For those following agent frameworks, this is a concrete prototype of "AI that improves AI," available for independent analysis via the open preprint. It is useful to compare the AIDE² loop — edit generation, benchmarking of its own versions, hidden evaluations — with familiar MLOps processes and conduct an audit: which improvements to your agents are currently done manually and could go through an automatic loop with independent evaluation. A separate design lesson: memory mechanisms that compress long context were derived by selection in this work, not designed manually. End users of AI services have nothing to deploy yet — there is no product; the conditional prospect is that if such loops are confirmed and become cheaper, the agents behind products may improve faster at the same development cost.

What Is Still Unknown / Limitations

There is no independent replication yet: according to available data, verification was limited to the internal self-consistency of the preprint, so the correct formulation is "verifiable," not "verified." External expertise has not begun to form: discussion on Hacker News has not started (1 point, 0 comments). AIDE² is not available as a product: the sources do not mention an API, prices, latency metrics, or the cost of an 8-day run; it is also unclear whether code and reproduction details will be published. It is unknown whether the gains will be reproduced on other agents and task sets and whether they will scale beyond the class of AI-R&D tasks; only three of the four held-out benchmarks are named. Finally, the reduction in reward hacking is currently an observation, not a proven mechanism.

Sources

Author

Look at AI, editorial team