How to use Unknown to improve a research agent
Weco used an autonomous outer research loop to rewrite and evaluate an inner agent harness, reporting seven improvements and testing selected versions on held-out tasks.
weco.aiPrompt to try
Editorial adaptation, not the source author’s exact prompt. Replace bracketed inputs and review tool permissions first.
Improve [research agent harness] through an outer loop that proposes code rewrites and evaluates them on [task suite] under [fixed cost budget]. Retain only improvements over the current best. Test selected versions on [held-out tasks], and report prompt size and reward-hacking behavior.
Reach people exploring AI agents
See placements →What happened
Weco reports seven improved versions over 100 iterations, held-out generalization, a 16-fold prompt reduction, and a reward-hacking rate reduction from 63% to 34% on a held-out GPU benchmark.
Reported by EvgeniyZh on BlogHow the agent did it
- Simplified AIDE by stripping ML-specific machinery.
- Ran an outer agent to propose inner-agent code rewrites.
- Evaluated proposals and retained only improvements.
Where a person stepped in
The team defined the evaluation setup and tested discovered agents on held-out tasks; the source does not describe reviewing every rewrite.
A person set the goal and reviewed the result.
Weco describes AIDE² as a system that runs an outer research loop to improve an inner research agent's harness. The outer loop uses the team's manually tuned AIDE human agent, powered by claude-opus-4.7. The inner agents use gemini-3-flash. These are different roles in the experiment, rather than one model doing every part of the work.
The team first simplified AIDE by removing ML-specific machinery, producing the starting inner agent AIDE 0. At each outer iteration, the system proposed a rewrite of the inner agent's code, evaluated the proposal across heterogeneous tasks, and kept it only if it improved on the previous best. The source reports 100 iterations, seven successive improved versions and eight unattended days.
The team tested agents discovered at steps 47 and 85 on held-out tasks. They report generalization beyond the optimization tasks and better results than their manually tuned baseline. The post also reports a 16-fold prompt-size reduction and defenses against reward hacking, with a held-out GPU benchmark rate falling from 63% to 34%. These are the authors' experimental claims, including their interpretation of Level 1 recursive self-improvement. Who it's for: research teams studying evaluation-driven improvements to agent harnesses.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Blog · EvgeniyZh
Frequently asked questions
How are proposed harness changes selected?
At each outer-loop step, the agent proposes an inner-agent code rewrite, evaluates it, and keeps it only if it is better than the previous best.
Which models power the two loops?
The supplied post says the outer AIDE human agent uses claude-opus-4.7, while each inner-loop agent runs on gemini-3-flash.
More details
Remaining steps
- Tested discovered agents on held-out tasks.
About this account
The outer agent autonomously rewrote and evaluated the inner harness during an eight-day unattended experiment.
Platform: Blog. Attribution: https://www.weco.ai/blog/first-evidence-of-recursive-self-improvement. Prompt: Suggested (written by Agents In Use, not the author's). Sourced by the Agents In Use pipeline.
Credit: EvgeniyZh on Blog
Source published 15 Jul 2026
Source checked 14 Oct 2026
Tools and topics
AIDE² · AIDE · claude-opus-4.7 · gemini-3-flash
Agent research · Evaluation · Self-improvement
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.