Evolutionary Search
Fitness supports selection; traces and critiques guide proposals. Evolutionary paradigms describe shared search structures across the reviewed methods.
Overview
Learning from experience, weights fixed
Execution experience can become persistent instructions, memory, workflows, and tools. This survey connects how those artifacts evolve with their evaluation, theoretical limits, deployment, and societal effects.
Read the abstract
Building agents that learn from experience is a central goal of artificial intelligence. Harness self-evolution offers a path toward this goal by revising an agent’s instructions, memory, workflows, and tools while model weights stay fixed. In this article, we develop a unified framework for how execution feedback becomes persistent harness change, covering systems that keep updating in operation and the offline optimizers they build on. We organize algorithms by their update targets, distinguishing instructions, memory, and workflow specifications interpreted by fixed components from executable tools and runtime logic, and compare how they propose and admit changes. Our theoretical analysis relates harness search to evolutionary algorithms and uses a continual PAC-learning perspective to give sufficient conditions under which an admission comparison certifies improvement. For evaluation, we distinguish four claims behind a reported gain: selection, whether the estimate is independent of the search or accounts for selection; activation, whether the accepted update operated in the scored runs; attribution, whether it caused the difference under controlled evaluation conditions; and improvement, whether the difference holds beyond the tasks that produced it. The reviewed reports often evaluate on held-out tasks, but evidence of activation and attribution remains limited. We then examine applications in personal assistance, enterprise automation, scientific discovery, robotics, and public services, and the implications of persistent, shareable harness state for work, markets, security, and governance. We conclude with open problems in optimization, infrastructure, evaluation, and responsible deployment. The project repository is available at https://github.com/zgcagi/Harness-RSI.
Figure 2. Research growth and representative systems, grouped by primary emphasis.
Definition
A persistent update loop
Execution feedback revises the harness \(H\) around a fixed model \(M\). Accepted changes persist across runs, and later experience can trigger further updates.
Self-evolution names the update mechanism; self-improvement requires a beneficial change supported by an explicit comparison. Harness RSI remains the conventional short name, without a guarantee that every accepted revision helps.
One update round
Run the current harness.
\[\tau_{k,i} \sim \operatorname{Run}(M, H_k, E_{k,i})\]
Collect a batch of trajectories.
\[\mathcal{B}_k = \{\tau_{k,i}\}_{i=1}^{b}\]
Use that feedback to revise the harness.
\[H_{k+1} = \operatorname{Update}(H_k, \mathcal{B}_k)\]
\(\tau\): trajectory · \(E\): environment · \(b\): batch size
The model and final evaluator stay fixed; internal critics may evolve. Rejected updates keep \(H_k\).
Taxonomy
What evolves
Two families of update targets: payloads interpreted by fixed host components, and executable implementations or compositions. Select a mechanism to browse its papers.
Figure 4. Harness update targets and the operational loop.
Full taxonomy and citation map (Figure 3)
Theory
Optimization and generalization
Two perspectives: how revisions are found, and what establishes generalization and retention.
Continual PAC Learning
An admission comparison can certify improvement when candidates are fixed before fresh validation tasks are drawn and error is controlled across comparisons. Repeated executions reduce run noise without adding independent tasks.
Search paradigms and representative methods
Elitist (1+1) / ES
SICAMCESkillOptSkillOpt-LiteSkillClaw
Elitist (1+λ)-style
Self-Harness
Archive-based program search
DGMDarwinXHGM
Recombination / crossover
DarwinXHarnessCompassSelf-HarnessMCETrace2Skill
Quality-diversity / MAP-Elites
HarnessBank
Cooperative coevolution
HarnessCompass
Multi-objective selection
Meta-HarnessSelf-HarnessSkillMOO
Bandit-guided evolution
HGM
Structural correspondences to search paradigms; they do not imply exact algorithm implementations.
Evaluation
Four claims behind a task score
Which evidence a study needs depends on its conclusion: comparing a final state differs from crediting an individual update and its lasting effect.
Condition
Selection
The reported estimate is unselected, or its inference accounts for selection. Uncorrected scores must remain separate from adaptive choices.
Step 1
Activation
Direct traces show the update operated in scored runs, with coverage recorded.
Step 2
Attribution
Controlled comparisons distinguish the update’s effect from execution variation, added inference compute, and evaluator change.
Step 3
Improvement
The difference holds away from the tasks that formed the update, with the baseline and round specified. The audit grades task distance only; it is not a numerical distance metric or proof of a reliable positive effect.
A held-out partition used for ranking, admission, or report choice is no longer independent of search. HarnessX checks replay, novelty, and structure without guaranteeing non-decreasing task scores; neither these checks nor Self-Harness's non-regression gate establishes retention outside the tested conditions.
Figure 5. Four claims and the controls needed across update rounds.
Evidence
What the evidence shows
95 methods in scope; 93 primary comparisons are scored. 2 methods without empirical evaluation are excluded. Criteria describe reported evidence within its recorded coverage.
Evidence by claim
Activation credits an identified scored-run case; Improvement grades task distance only.
Criterion metPartialNot met
Criteria met
Coverage of the four criteria; not an overall method ranking.
Activation and attribution remain the main evidence gaps. A held-out gain at one round also leaves long-term retention unresolved.
Applications & outlook
Applications and open questions
Persistent harness state can carry personal preferences, shared enterprise procedures, research know-how, and robotic skills. The survey examines existing and prospective deployments, alongside effects on work, knowledge ownership, security, and governance.
Open questions span model–harness co-evolution, optimization, state infrastructure, dynamic evaluation, safety boundaries, and provenance. Reusable experience must remain useful as tasks, models, and deployments change.
Applications and societal implications (Figures 7–8)
Figure 7. Deployment requirements across five settings.
Figure 8. Hypothesized pathways, without established causal estimates.
Reading list
Browse the list
Search papers and mechanisms. Evidence markers follow Selection · Activation · Attribution · Improvement.
Citation
Cite this work
BibTeX for the current working draft.
BibTeX
@misc{wei2026harnessrsi, title = {{Harness Self-Evolution: Theory, Algorithms, Evaluation, Applications, and Impact}}, author = {Wei, Chuyang and Guan, Haoxiang and Zhu, Minghang and Zhou, Kexin and He, Jiyan and Zhou, Xiren and Chen, Huanhuan and Shen, Yifei}, year = {2026}, note = {Working draft, 12 October 2026}, url = {https://openreview.net/pdf?id=m5jsFGXUM2} }