AI agent improved its own code for 8 days

A paper on arXiv (2609.26457) describes the AIDE² system, which found seven improvements to its own code during 8 days of autonomous operation, showing gains on four held-out benchmarks — from ML engineering to out-of-distribution weather forecasting.

Compiled by Sergey KostenchukPublished 2026-09-23Updated 2026-09-24

2026-09-23 Research

🤖 AI agent improved its own code for 8 days

The AIDE² system proposes edits to its own code and keeps only those that perform better on hidden evaluations of AI-R&D tasks. Over 8 days of autonomous operation, seven improvements were found — from a new search policy to memory that compresses growing context.

🌍 Gains transferred to four held-out benchmarks: ML engineering, heuristic algorithms, and out-of-distribution weather forecasting. On all of them, the best agent matches or outperforms the production agent from the strongest on FML-Bench. Bonus: reward hacking dropped from 55% to 32%, although the loop did not explicitly optimize for this.

👤 First verifiable closed loop of self-improvement: “verifiable,” not “verified” — no replications yet. Those following agent frameworks should compare AIDE² with human MLOps processes.

Source: https://arxiv.org/abs/2609.26457

Sources