PaperlayerView exact revision on arXiv

Source-linked AI summary

FreeEvolve: Learning to Evolve Beyond Fixed Loops

Lecheng Kong, Like Hui, Nikos Kanakaris, Prithwish Jana, Sahika Genc, Narayanan Sadagopan

arXiv:2610.09197v1cs.AIcs.LG

TL;DR

Agent evolvers automate agent-harness design, but their optimization loops remain hand-designed. FREEEVOLVE lets an evolver control a bounded improvement campaign and learns how to organize it through meta-evolution. Across four environments, it matches or exceeds hand-designed evolvers on six of seven held-out metrics.

  • Problem

    Agent evolvers automate harness design, but a hand-designed loop still fixes evaluation, candidate comparison, and stopping decisions before evolution begins.

    Source: §1 · p. 1

  • Method

    FREEEVOLVE gives the evolver control of a bounded improvement campaign, using evidence from each execution to choose what to do next.

    Source: §1 · p. 2, §3.1 · p. 4

  • Results

    Across four environments, meta-evolution raises held-out mean performance from 33.3% to 51.7% on τ 3-bench, 55.4% to 69.1% on ARC-AGI-2, 61.0% to 81.6% on ARC-AGI-3, and 80.9% to 82.5% on Terminal-Bench 2.1, matching or exceeding hand-designed evolvers on six of seven metrics.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The learned evolution process improves with experience and transfers as a way of organizing campaigns rather than as domain knowledge.

    Source: §1 · p. 3, §5 · p. 10

    FREEEVOLVE does not make autonomous improvement safe for high-stakes deployment, where domain-specific oversight remains necessary.

    Source: ETHICS STATEMENT · p. 11

Abstract

from arXiv · show

Agent evolvers automate the design of the prompts, skills and workflows around language model agents, yet the optimization process they follow is still designed by hand: a fixed search loop decides how candidates are evaluated, which are kept and when the search stops. We propose FREEEVOLVE, which automates this process as well. An environment specifies the goal, target agent, evaluator, data and resource limits; within these limits, the evolver itself decides what to test, how much evidence to collect, which candidates to pursue and when to stop. These decisions follow an editable evolution skill, which we improve through meta-evolution by scoring each candidate skill on the fresh target agent it produces. The optimization process thus becomes a capability learned from experience rather than a loop engineered in advance. On tau3-bench, ARC-AGI-2, ARC-AGI-3 and Terminal-Bench 2.1, FREEEVOLVE controls the evolution campaign by itself, yet improves the primary held-out metric by 13.6 points on average and matches or exceeds hand-designed evolvers. The learned process keeps improving with experience: meta-evolved skills add 6.9 points over the seed skill on fresh target agents, demonstrating transferability across environments.

1 INTRODUCTION

Agent evolvers improve harnesses, but humans still prescribe the search loop that determines what to evaluate, keep, and when to stop. FREEEVOLVE gives campaign decisions to an evolver operating within a bounded environment, then improves its evolution skill through complete campaigns; across four environments, it improves held-out performance and transfers learned skills.

Motivation

  • The paper asks whether evolvers can organize their own improvement campaigns and learn how to do so, rather than relying on human-designed loops that fix evaluation and stopping decisions in advance.Such loops encode assumptions about noise, task difficulty, curricula, and parallelism that may change across agents and environments.Source: §1 · p. 1

Automating the optimization process

  • FREEEVOLVE treats campaign organization as a capability: a bounded environment specifies the goal, editable target, available data, and resource constraints, while the evolver decides what to do as evidence arrives.Campaign decisions include task focus, repeated comparisons, candidate exploration, resource allocation, and stopping.Source: §1 · pp. 2–3

Learning the process

  • The campaign policy is an editable evolution skill; meta-evolution evaluates candidate skills through complete improvement campaigns and refines them using resulting agents and campaign traces.The resulting skill can be applied unchanged in new environments without retaining the original agents or execution artifacts.Source: §1 · pp. 2–3

Experimental results

  • Held-out mean performance rises from 33.3% to 51.7% on τ 3-bench, 55.4% to 69.1% on ARC-AGI-2, 61.0% to 81.6% on ARC-AGI-3, and 80.9% to 82.5% on Terminal-Bench 2.1.FREEEVOLVE matches or exceeds hand-designed evolvers on six of seven held-out metrics.Source: §1 · pp. 2–3
  • A skill meta-evolved on τ 3-bench reaches 64.3% mean accuracy on ARC-AGI-2 and 71.0% RHAE on ARC-AGI-3, matching or exceeding skills meta-evolved directly there.The learned skill outperforms the seed skill on all four mean metrics and all three available pass@3 metrics on fresh workers.Source: §1 · pp. 2–3

2 RELATED WORK

Prior work automates harness components and, in some cases, the mechanism that improves them, but typically leaves the campaign itself fixed. FREEEVOLVE instead learns a reusable way to organize improvement within a run.

Automating harness design

  • Systems that automate harness design edit prompts, skills, code, or workflows, while the search process deciding evaluations, candidate retention, and stopping is typically specified in advance.Evolutionary coding systems also improve programs and agent harnesses through model-generated edits.Source: §2 · p. 3

Automating the improver

  • Meta-evolution and self-improvement systems make improvement mechanisms editable, but prior work focuses on improving a fixed campaign.HyperAgents jointly evolves task agents and meta-agents, allowing its procedure to change across outer iterations.Source: §2 · p. 3

Automating the improvement process

  • FREEEVOLVE extends automation to campaign decisions within a run and learns a reusable way of organizing improvement from completed campaigns.The learned process is neither a better artifact nor a search procedure for the next run.Source: §2 · p. 3

3 METHOD

FREEEVOLVE gives the evolver control of the improvement campaign inside a bounded environment, then meta-evolves the skill that organizes those campaigns. The same environment interface and evolution operation support both levels.

3 METHOD

  • FREEEVOLVE uses the same evolution operation at two levels: an evolver organizes agent campaigns, while a meta-evolver improves the evolution skill.The bounded environment specifies the goal, editable target, available data, and resource constraints.Source: §3 · p. 3

3.1 A BOUNDED INTERFACE FOR CAMPAIGN AUTONOMY

  • The bounded environment fixes the goal, data, editable target, evaluation rules, and resource limits, while valid requests leave execution choices to the evolver.For benchmark workers, the evolver can select tasks, repetitions and seeds, concurrency, and timeouts subject to benchmark rules and resource bounds.Source: §3.1 · p. 4

3.2 THE EVOLVER CONTROLS THE CAMPAIGN

  • The framework invokes the evolver once; it selects candidates and requests, uses returned artifacts to guide subsequent decisions, and decides when to stop.The seed skill supplies general dispositions, but leaves task selection, repetitions, branching, and acceptance to the evolver.Source: §3.2 · p. 4

3.3 META-EVOLUTION LEARNS THE CAMPAIGN POLICY

  • Unlike a procedure fixed during each inner run, the meta-evolved skill stays active and organizes evidence gathering and search as results arrive.The skill is the editable target of meta-evolution.Source: §3.3 · p. 4
  • Meta-evolution judges candidate skills through complete worker campaigns and their resulting workers and traces, with the outer evolver using campaign outcomes as feedback.The outer controller and candidate skill begin as separate copies, so editing the candidate does not change the controller.Source: §3.3 · p. 5, Figure 2 · p. 5
  • The environment–evolver boundary supports Slurm execution while leaving the evolver, evolution skill, and Evolve operation unchanged.The environment-specific execution function launches inner campaigns and returns artifacts as they finish; its constraints expose node allocation and concurrency limits.Source: §3.3 · p. 5
  • A meta-evolved skill can be frozen and reused across environments with different tasks, models, or hardware without carrying over source workers or campaign state.Source: §3.3 · p. 5
  • Table 1 reports that FREEEVOLVE matches or exceeds hand-designed evolvers on held-out worker performance.The table compares the held-out performance of the worker returned by each method.Source: Table 1 · p. 6

4 EXPERIMENTS

The experiments test whether FREEEVOLVE can produce strong held-out workers, learn a stronger evolution skill, transfer that skill across environments and models, and adapt its campaign decisions. Across four environments, it improves the primary held-out metric by 13.6 points on average, while further experiments examine skill transfer, model capability, and campaign control.

4.1 EXPERIMENTAL SETUP

  • The evaluation covers four environments spanning conversational tool use, abstract reasoning, interactive games, and terminal control, measuring unseen-task performance after search on an evolution partition.Held-out workers are evaluated three times; the paper reports mean and standard deviation and, where defined, pass@3.Source: §4.1 · p. 6

4.2 LEARNED AUTONOMOUS CAMPAIGNS PRODUCE STRONG WORKERS

  • 13.6 points on average: FREEEVOLVE improves the primary held-out metric across four environments, outperforming expert-designed evolvers on three benchmarks and matching or exceeding EvoX and HyperAgents.The comparison evaluates the methods under their own campaign organizations.Source: §4.2 · p. 7

4.3 META-EVOLUTION PRODUCES A STRONGER EVOLVER

  • 6.9 points on average: the meta-evolved skill improves over the seed across four mean metrics, improves every available pass@3 metric, and beats the no-data rewrite in every environment.Each skill is tested on the same fresh worker with the same budget, without carrying over an evolved worker or campaign artifact.Source: §4.3 · p. 7

4.4 LEARNED EVOLUTION SKILLS TRANSFER ACROSS ENVIRONMENTS AND MODELS

  • Single-source skills transfer unevenly and can reduce Terminal-Bench 2.1 mean, while joint meta-evolution on τ 3-bench and Terminal-Bench 2.1 improves every reported target metric over the seed.The transfer test freezes each skill and installs it on a fresh target worker.Source: §4.4 · p. 7, Table 2 · p. 8
  • 5.3 points: FREEEVOLVE’s frozen harness retains this average gain over the base harness across four non-native worker-model settings and is strongest in all six settings.Figure 4 also compares ARC-AGI-3 training RHAE against total tokens used to produce each skill; by 0.74B tokens, FREEEVOLVE reaches a score no baseline attains with more than twice the token cost.Source: Figure 4 · p. 8, §4.4 · p. 8

4.5 MODEL CAPABILITY SUPPORTS EFFECTIVE CAMPAIGN CONTROL

  • Held-out improvement begins with Sonnet 4.6 and generally increases with newer and stronger evolver models, although Sonnet 4.6 outperforms Opus 4.6.Haiku 4.5 does not improve either held-out metric; the authors identify sufficient model capability as a limitation and leave broader study of the boundary to future work.Source: §4.5 · pp. 8–9

4.6 AUTONOMOUS CAMPAIGNS ADAPT EVALUATION AND SEARCH

  • Always evaluating the full task set reduces the decisions possible within the time budget, while fixed concurrency and no timeouts can waste tokens on slow or failed executions and yield a weaker worker.The interventions hold the learned skill and resource budget fixed while removing control over one environment-facing decision at a time.Source: §4.6 · p. 9, Table 4 · p. 9
  • In a traced ARC-AGI-2 campaign, the evolver adapts evaluation resolution, branches search, combines discoveries, and selects a worker as evidence accumulates; learned skills operationalize these choices through task selection, evaluation allocation, and search control.The figure annotates decisions such as repeating a promising version, undoing a regression, and branching an ancestor into four variants.Source: §4.6 · p. 9, Figure 5 · p. 9

5 CONCLUSION

FREEEVOLVE automates campaign organization and learns how to run improvement campaigns from their outcomes. Across four environments, its self-controlled process matches or exceeds hand-designed evolvers, while carrying campaign procedure rather than domain knowledge.

5 CONCLUSION

  • Across four environments, FREEEVOLVE’s self-controlled campaign process matches or exceeds hand-designed evolvers.The learned process improves with experience and carries a way of running campaigns rather than domain knowledge.Source: §5 · p. 10
  • The study uses Claude models for evolvers, workers, simulated users, and judges; authors designed experiments, checked reported numbers, and take responsibility for the content.Runs preserve the installed environment, evaluated target, evolver instructions, execution trace, and raw evaluation results.Source: AI USE STATEMENT · p. 11, REPRODUCIBILITY STATEMENT · p. 11
  • Meta-evolving harnesses can amplify errors, consume substantial compute, and obscure the source of behavioral changes, so high-stakes deployment still needs domain-specific oversight.Nesting limits, retained evaluation records, fixed model capability, and separation of candidate generation from acceptance improve auditability but do not establish safety.Source: ETHICS STATEMENT · p. 11

A IMPLEMENTATION DETAILS

Implementation uses a single evolver invocation within a fixed environment boundary, leaving campaign scheduling to the evolver. The same interface supports bounded meta-evolution, with explicit standalone and outer-campaign resource limits.

A IMPLEMENTATION DETAILS

  • The seed skill gives general evidence-based instructions but prescribes no iteration count, branching, curriculum, evaluation schedule, or stopping rule; one invocation runs without a framework-level loop.A single evolve.sh call installs the environment and invokes Claude Code with SKILL.md.Source: A IMPLEMENTATION DETAILS · p. 14
  • Each environment specifies its goal, data, constraints, and artifacts; .target declares the editable surface, run.sh implements RE, and the workspace stores campaign outputs.Environment files remain fixed during a campaign, while only the declared target is editable.Source: A IMPLEMENTATION DETAILS · p. 14
  • RE(x, u) installs candidate x at the declared path and passes u through run.sh; recorded measurements and traces evaluate the same declared edit surface.The complete run directory lets the evolver invoke runs in any order, branch independently, and decide when evidence is sufficient, without a framework search schedule.Source: A IMPLEMENTATION DETAILS · p. 14
  • The meta-environment runs a candidate SKILL.md as an evolver in a fresh worker and returns the full inner run directory; meta-environments are excluded from the catalogue, bounding execution to two levels.The outer evolver chooses the worker environment and number of inner runs.Source: A IMPLEMENTATION DETAILS · p. 14
  • Standalone worker runs have a 12-hour budget on one 128-CPU machine and at most 24 concurrent task attempts; meta-evolution has a 24-hour outer budget and at most five simultaneous inner campaigns.Inner evolver runtime within meta-evolution has no separate wall-clock limit.Source: A IMPLEMENTATION DETAILS · p. 14

B DATASET DETAILS

The dataset appendix describes four benchmark settings, their held-out evaluation measures, and the separation of evolution and test partitions. The benchmarks span tool-mediated banking, grid reasoning, interactive games, and command-line tasks.

B DATASET DETAILS

  • τ3-bench uses simulated-customer banking conversations graded with binary rewards, with 68 evolution tasks and 29 held-out tasks from 97 total.Agents retrieve bank policy, take required actions, and communicate a resolution.Source: B DATASET DETAILS · p. 14
  • ARC-AGI-2 evaluates few-shot grid transformations using 56 evolution tasks and a disjoint set of 111 held-out tasks, with held-out task accuracy reported.Meta-evolution may use only the 56-task evolution partition.Source: B DATASET DETAILS · pp. 14–15
  • Table 5 covers benchmark datasets, metrics, evaluation protocols, and base agents.Source: Table 5 · p. 15
  • ARC-AGI-3 uses 17 evolution games and 8 disjoint held-out games, reporting RHAE as a percentage based on level completion and action efficiency.Later levels receive greater linear weight; uncompleted levels score zero and faster-than-human play is capped.Source: B DATASET DETAILS · p. 15
  • Terminal-Bench 2.1 reports mean and pass@3 on 61 held-out command-line tasks verified by test suites in isolated containers.Tasks cover software debugging, systems configuration, security, data processing, and scientific workflows.Source: B DATASET DETAILS · p. 15
  • All partitions are fixed before meta-evolution, held-out data is unavailable to both evolver levels, and the same split is used across generations within each benchmark.The authors searched run directories for held-out data IDs and report that evolvers stayed within the established directory.Source: B DATASET DETAILS · p. 15

C CONFIGURATION DETAILS

Configurations hold model capability and environment boundaries fixed while allowing edits to declared agent harnesses and search procedures. The appendix details benchmark-specific editable surfaces and identifies the scope of campaign-rule and fresh-worker comparisons.

C CONFIGURATION DETAILS

  • Worker, worker-level evolver, and outer meta-evolver use the same Claude model within each environment, and the evolver cannot alter the environment or redirect evaluation from the declared target.The model-capability experiment is the stated exception to model matching.Source: C CONFIGURATION DETAILS · p. 15
  • For τ3-bench, the base is the vendored customer-service agent; the editable surface spans its prompt, tool-use loop, response parsing, retries, and termination logic, while model capability and task resources remain out of scope.The agent, simulated user, and judge use the specified Claude Opus 4.8 model, with reasoning effort fixed to medium.Source: C CONFIGURATION DETAILS · pp. 15–16
  • Figure 6 identifies campaign rules added to the seed skill by meta-evolution; Table 6 scopes a fresh-worker comparison of seed, no-data, and meta-evolved evolvers.The table headings identify common seed disposition, learned policy, and realized advantage.Source: Figure 6 · p. 16, C CONFIGURATION DETAILS · p. 16, Table 6 · p. 16
  • ARC-AGI-2 holds model and inference configuration fixed across conditions and evolves the solver harness and search procedure, not grids, splits, or evaluator.Each worker is produced from the same base solver rather than inherited from the prior condition.Source: C CONFIGURATION DETAILS · p. 16
  • ARC-AGI-3 fixes the model, reasoning configuration, games, and evaluator while permitting changes to the declared harness and search procedure.Source: C CONFIGURATION DETAILS · p. 16
  • Terminal-Bench 2.1 permits edits to the control loop, prompts, parsers, and terminal-session handling while keeping containers, task statements, test suites, and verifier code fixed.The base worker is the unmodified deep agent using Claude Opus 4.8 at temperature 1.0.Source: C CONFIGURATION DETAILS · p. 16

D ADDITIONAL RESULTS

Evolver-controlled campaigns improve the complete system across all four environments. Additional results report fresh-worker and model-capability comparisons, fixed-decision ablations, and campaign wall-clock times.

E ADDITIONAL RELATED WORK

Prior work evolves harnesses, workflows, or self-improving systems, and some studies test whether evolved products transfer. FREEEVOLVE instead tests whether a learned campaign-organizing skill transfers to a fresh worker.

  • FREEEVOLVE differs from harness-evolution work by learning a reusable skill that organizes future improvement campaigns rather than only searching or adapting harness components.Source: E ADDITIONAL RELATED WORK · p. 17
  • Several cited studies test transfer of evolved harnesses or agents, while FREEEVOLVE freezes its evolution skill before a fresh worker sees the target environment.Source: E ADDITIONAL RELATED WORK · pp. 17–18
  • Table 10 reports complete cross-model transfer results for frozen worker harnesses.Source: Table 10 · p. 18
  • Table 11 maps questions Q1–Q5 to the main-text evidence that addresses them.Source: Table 11 · p. 18

F CONCRETE EXAMPLE FOR WORKER AND META-EVOLVER ENVIRONMENTS

Worker and meta-evolver environments share a contract format but differ in their targets and evaluation units. An ARC-AGI-3 example traces how a candidate skill is tested through a complete worker-evolution run.

G PROMPTS

The prompts describe a canonical seed skill and a more detailed meta-evolved procedure. They specify how to evaluate targets, use evidence, and hand off the best-supported version.

G.1 CANONICAL SEED EVOLUTION SKILL

  • Every reported meta-evolution lineage starts from the same canonical seed, which is task-agnostic.Source: §G.1 · p. 20
  • The seed directs the evolver to read ENVIRONMENT.md as the source of truth and edit the named agent in place for evaluation.Source: §G.1 · p. 20
  • The seed instructs the evolver to spend the available wall-clock budget, leave the strongest version in best/, and summarize its evidence in RESULT.md.Source: §G.1 · p. 20
  • Table 13 contrasts choices fixed by the canonical seed with those available for meta-evolution to learn.Source: Table 13 · p. 20
  • The observed evolver considers evaluation counts and paired comparisons when selecting the delivered harness, not benchmark scores alone.Source: §G.1 · p. 20

G.2 META-EVOLVED INSTRUCTION

  • The meta-evolved procedure adds baseline establishment, failure diagnosis, varied changes, noisy-measurement handling, candidate retention, and restoration of the best-supported target.Source: §G.2 · p. 20
  • The evolved instruction treats raw measurements and traces as evidence about the learner and distinguishes the best-supported version from the last edited state.Source: §G.2 · p. 20
Loading 2610.09197v1…