Skip to content
AI-generated research idea🔴Hard🔬Significant ResearchClosely matching work found

Meta-Recursive Co-Evolution of Improver Stage Instructions and Composition Rules

Co-evolve the meta-prompt rules and composition heuristics using retrospective edit-attribution logs so an agent improves how it edits, not just its stored examples.

Time

a couple of months

Compute level

Single GPU or less

Resources

Workstation with Docker container runtime, substantial API budget for Claude Opus 4.8 and Gemini 3.5 Flash/GLM-5 across replay buffers and test streams, and WorkBuddy-Bench, JobBench, and LAB.

What the paper used

Frozen API executors (Gemini 3.5 Flash, GLM-5) and editor/selector (Claude Opus 4.8 at high reasoning effort) run inside OpenCode/Codex Docker containers (200-turn and 3,600s rollout caps), spending roughly 1-100M tokens per task across WorkBuddy-Bench (100 test tasks), JobBench (34 test tasks), and LAB (100 test tasks) over 3 runs.

Prerequisites

Recursive self-improvement, meta-prompt optimization, and LLM agent architectures.

Scope

new model architecture

The Gap

In Section 2.1, Section 2.3, and Appendix H, harnessing-level recursion updates only the improvement library (adopted harnesses and EXPERIENCE.md entries), while the editor, composer, selector, and reflector prompts (Prompts 1-6) and decision rules remain completely static.

“Model weights, stage instructions, and the decision rule remain fixed, while the library evolves.”
From the paper, Section 2.3, page 7. Matched against the PDF text.

Limitation

Because the improvement algorithm and stage instructions never adapt, systematic meta-level weaknesses—such as over-prescriptive guardrails, selector precision bias, or suboptimal component choices across different executor models—persist unchanged across the entire deployment stream.

Proposed Solution

Add an outer meta-update loop on top of Experience Reflector logs: after every M tasks, run a Meta-Reflector over the accumulated edit-attribution rows (helped/hurt/confounded verdicts across components C1-C8) to diagnose systematic improver failures (e.g., high hurt rates on guardrails vs. skills, or frequent validation fallbacks). Propose targeted diffs to modular strategy blocks within the Editor, Composer, and Selector prompts, and validate candidate meta-prompt variants offline by replaying a small buffer of past contrasting trajectory pairs (verifying whether the updated meta-prompt avoids past hurt edits while preserving helped edits) before promoting the updated improver harness.

Why This Matters

Enabling the improver's own policies to adapt could yield true cumulative self-improvement in editing quality across shifting domains and executor models.

How to Get Started

How to test it: Test whether periodic meta-reflection over Experience Reflector attribution logs improves downstream harness-editing efficiency and cross-benchmark transfer compared to static Prompts 1-6. Evaluate on a held-out stream (e.g., meta-adapting on JobBench and testing on WorkBuddy-Bench) measuring final task scores and edit hurt-to-help ratios; stop if evolved stage instructions induce instability or higher validation-failure rates.

  1. Factor Prompts 1-6 in github.com/google-research/harness2 into immutable safety/format invariants and an editable strategy-rules file.
  2. Run a baseline stream on JobBench and export all Experience Reflector edit-attribution rows (component, verdict, observed contrast).
  3. Design a Meta-Reflector module that proposes rule edits from aggregated hurt/confounded patterns and scores them on how accurately they predict or revise historical helped vs. hurt decisions.
  4. Evaluate the meta-updated improver on cross-benchmark transfer (JobBench to WorkBuddy-Bench) against the static-prompt Harness^2 baseline.

Suggested by AI. Adapt it to your setup.

Has This Been Done?

Closely matching work found

The literature provided in [0], [1], [2], and [6] extensively covers the concept of recursive self-improvement and meta-recursive evolution of harness instructions and components. The proposal essentially describes the cycle of meta-reflection and validation already standard in these RSI (Recursive Self-Improvement) frameworks.

We searched OpenAlex and arXiv for “meta-recursive self-improvement agent editable improvement strategy harness” and “Meta-Recursive Co-Evolution of Improver Stage Instructions and Composition Rules” and an AI compared the idea with the top 10 results. A small search can miss relevant work.

Source Paper

Harness^2: Recursive Agent Harnessing for an Open World

Ruiyao Xu, Yanfei Chen, Zhongying CuiZhu +6 more · 2026

This paper introduces Harness^2, a plug-and-play recursive agent harnessing framework that enables LLM agents to optimize their execution harnesses at test time in open-world settings without ground-truth verifiers. It couples task-level contrastive refinement (Propose-Probe-Compose) with harnessing-level recursion that distills reusable harness components and lessons into an evolving improvement library across tasks.

Related Ideas

Is this a good idea?

Share onXLinkedIn