Alphabell.
Paper · Self-modifying agents

AIDE2

Recursive self-improvement of AI research agents (AIDE2)AIDE2 is a two-level system in which a research agent rewrites the code of an AIDE-style ML research agent, grades each rewrite on hidden held-out scores across AI R&D tasks, and keeps only rewrites that beat the incumbent. The paper reports an 8-day autonomous run as evidence of recursive self-improvement at the harness layer.

The loop

An outer-loop research agent (Weco's production AIDE agent on Claude Opus 4.7) proposes rewrites to the code of an inner research agent (starting from AIDE0 on Gemini 3 Flash), including its search policy and its memory and context management. Each rewrite is graded by running it on ML engineering, heuristic-algorithm and harness-engineering tasks scored on private held-out data, and an accepted rewrite becomes the incumbent that the next step edits. In an "ignition test" a discovered agent (AIDE47) was itself placed in the outer-loop seat and kept producing accepted rewrites, though the authors call that comparison inconclusive.

The loop
AIDE2
Paper · arXiv
  1. Outer agent rewrites inner agent code
  2. Accepted rewrites become the incumbent agent
  3. Discovered agent becomes the outer loop agent
  1. Outer agent rewrites inner agent code
  2. Accepted rewrites become the incumbent agent
  3. Discovered agent becomes the outer loop agent
↻ The improved system does the next round, and the loop turns again.

Why it is a road to recursion

The object being optimized is the AI researcher itself, so each accepted rewrite is a better research agent that, as the ignition test probes, can take over the job of improving the next version.

Evidence

In an autonomous 8-day run the loop accepted 7 of 99 proposed rewrites (incumbent grade 0.703 -> 0.778), the strongest discovered agent matched or beat Weco's human-engineered production agent on four held-out benchmarks (ALE-Bench, MLE-Bench, FML-Bench, WeatherBench 2), and reward hacking on a separate held-out task family fell from 55% to 32%.

recursive-self-improvementresearch-agentsharnessheld-out-evaluationreward-hacking

Related loops

More self-modifying agents →
PaperarXiv
Darwin Gödel Machine
  1. Coding agent edits its own code
  2. Scored child agents enter the archive
  3. Archived agents make later self modifications
  4. repeat
Self-modifying agentsUBC, Vector Institute, Sakana AI · 2025
ProjectAnthropic
Claude Code
  1. Claude Code writes its own code changes
  2. Changes become next Claude Code versions
  3. Release becomes harness for next development round
  4. repeat
Self-modifying agentsAnthropic · 2025
PaperarXiv
SICA
  1. Best agent changes its own code
  2. Edited agent is benchmarked and archived
  3. Archived agent makes the next edit
  4. repeat
Self-modifying agentsUniversity of Bristol, iGent AI · 2025