RRSI: Regularized Recursive Self-Improvement of Agent Harnesses
Regularize the search trajectory used to evolve an agent harness.
Key takeaway
RRSI adds proposal-side sparsity, evidence-aware exploration, leakage screening, noise-aware acceptance, cost-aware acceptance and pruning; it improves held-out performance and reduces policy-token use.
Updated on Sep 23, 2026
01
Abstract
RRSI regularizes recursive self-improvement of agent harnesses by constraining both proposed edits and the changes allowed to persist. The goal is to improve held-out performance without allowing harness evolution to overfit benchmarks, chase noise, or accumulate unnecessary complexity.
02
Introduction
Agent behavior depends on prompts, control flow, tools, memory, and context management. Manual tuning is slow, while unrestricted automated evolution can optimize the evaluation loop instead of the underlying capability.
03
Problem
Harness evolution must balance performance gains against benchmark leakage, stochastic evaluation noise, policy-token cost, and growing system complexity.
04
Background
RRSI connects recursive self-improvement with automated harness evolution, meta-learning, test-time optimization, and regularized model-selection ideas.
05
Methodology
The system evaluates the current harness, summarizes feedback, proposes candidates, screens for leakage, evaluates candidates, applies noise and cost gates, prunes unproductive components, and finally evaluates the selected harness on held-out benchmarks.
Formula notes
Mathematical details
Score and cost
Expected task score and policy-token cost.
- A
- -agent
- D
- -task set
- H
- -harness
- c
- -policy-token cost
- r
- -reward
- tau
- -trajectory
Annealed edit budget
Attributable edit budget decreases over rounds.
- T
- -total rounds
- t
- -round
- b_max
- -maximum
- b_min
- -minimum
Cost-aware acceptance
Additional cost must be justified by score gain.
- beta0
- -base tolerance
- beta1
- -cost per gain
- Delta C
- -relative cost change
- Delta S
- -score gain
Noise-adjusted floor
Candidate must clear the empirical noise floor.
- delta
- -noise band
- S_star
- -best evolve score
06
Architecture
The architecture consists of a frozen backbone plus editable harness, proposer and analyst components, leakage critic, selection gates, cost controls, and held-out evaluation. Proposal-side and selection-side regularization constrain different failure modes.
07
Dataset
The evaluation spans Terminal-Bench 2.1, SWE-bench Verified, Harvey LAB, JobBench, GDPval, APEX-Agents, EngDesign, and Frontier-Eng.
08
Training
The backbone remains frozen. Improvement occurs through recursive modification of the harness and test-time evaluation rather than conventional model-weight training.
09
Experiments
Reported gains include +6.0 on Terminal-Bench, +4.9 on EngDesign, +1.1 on Harvey Evolve, +1.8 on SWE-bench Verified, +2.3 on Harvey held-out, +3.5–4.7 on JobBench/GDPval/APEX, and +4.3 Medal points on Frontier-Eng.
Evidence highlights
- 01Production
- 02Trade-offs
- 03Failure modes
- 04Scaling
10
Baselines
Comparisons include the initial H0 harness and prior harness-evolution methods such as Meta-Harness, AHE, TTHE, and HarnessX.
11
Results
RRSI improves held-out splits with reported gains up to 4.7 points OOD. Reported token savings differ between the abstract and project summary and should remain attributed.
12
Ablation
Proposal-side controls, selection-side controls, leakage screening, cost gates, and pruning address distinct failure modes.
13
Limitations
Results remain sensitive to stochasticity, leakage, selection noise, held-out split quality, and the cost of candidate evaluation.
14
Conclusion
RRSI treats harness evolution as regularized optimization, constraining leakage, noise, cost, and unnecessary complexity while preserving useful changes.
15
References
RRSI primary paper and related work on recursive self-improvement, automated harness evolution, Meta-Harness, AHE, TTHE, and HarnessX.
Source trail
Primary papers, implementation pages, and external source records used to ground this explainer.
Continue reading
Related research
GitSwarm
GitSwarm turns inference-time work into a persistent, branchable body of reusable contributions that later agents can build on.
AI AgentsLongHarness
LongHarness helps agents reason over 100M-token document collections by keeping evidence and conclusions as explicit, revisable states that guide one another.
AI AgentsGenerative Agents
An agent architecture combining memory, retrieval, reflection, and planning to simulate believable long-horizon behavior.