Harnessing Agentic Evolution — part 1/3

Cleaned-text code points [0, 15937). Bibliography omitted here; source.md includes it.

# Harnessing Agentic Evolution

###### Abstract

Agentic evolution has emerged as a powerful paradigm for improving programs, workflows, and scientific solutions by iteratively generating candidates, evaluating them, and using feedback to guide future search. However, existing methods are typically instantiated either as fixed hand-designed procedures that are modular but rigid, or as general-purpose agents that flexibly integrate feedback but can drift in long-horizon evolution. Both forms accumulate rich evidence over time, including candidates, feedback, traces, and failures, yet lack a stable interface for organizing this evidence and revising the mechanism that drives future evolution. We address this limitation by formulating agentic evolution as an interactive environment , where the accumulated evolution context serves as a process-level state. We introduce AEvo , a harnessed meta-editing framework in which a meta-agent observes this state and acts not by directly proposing the next candidate, but by editing the procedure or agent context that controls future evolution. This unified interface enables AEvo to steer both procedure-based and agent-based evolution, making accumulated evidence actionable for long-horizon search. Empirical evaluations on agentic and reasoning benchmarks show that AEvo outperforms five evolution baselines, achieving a 26% relative improvement over the strongest baseline. Across three open-ended optimization tasks, AEvo further outperforms four evolution baselines and achieves state-of-the-art performance under the same iteration budget.

## 1 Introduction

Agentic evolution reframes LLM-based problem solving as a process of constructing and revising solutions [ 16 , 7 ] . Instead of treating the model only as a generator of candidate answers, these methods use LLMs, agentic workflows, or coding agents to drive iterative improvement: produce candidate artifacts, interpret feedback from evaluation, and influence what the system explores next [ 22 ] . This paradigm has been applied to program synthesis [ 10 ] , scientific discovery [ 17 , 38 ] , systems optimization [ 25 , 6 ] , and agent self-improvement [ 41 , 27 , 23 ] . In this paper, we use agentic evolution to broadly refer to evolution processes whose search behavior is driven by either structured agentic procedures or general-purpose agents.

Existing agentic evolution methods typically instantiate this paradigm in two ways. In procedure-based evolution , a predefined outer loop controls parent selection, candidate generation, evaluation, and population update [ 18 , 36 , 42 ] . This makes evolution modular and reproducible, but also ties long-horizon search to fixed selection rules, feedback summaries, and update heuristics. In agent-based evolution , a general-purpose agent manages the search process by observing feedback, inspecting traces, editing candidates, writing tools, and deciding what to try next [ 11 , 22 ] . This gives evolution greater flexibility, but the agent can drift as candidates, logs, hypotheses, and intermediate files accumulate. In both cases, long-horizon evolution remains prone to local optima: procedures may repeatedly exploit the same hand-designed search pattern, while agents may overcommit to misleading evidence or stale assumptions in a growing context.

Recent work has tried to address these limitations by either broadening exploration with collaborative agents [ 22 ] or making the evolution mechanism self-modifying [ 41 ] . These directions show that stronger search context and editable improvement mechanisms are useful, but they do not by themselves provide a stable interface for long-horizon evolution. The core challenge is that evolution accumulates candidates, feedback, traces, failures, and intermediate decisions over time, yet lacks a unified way to organize this evidence and revise the mechanism that drives future evolution.

We address this challenge by formulating agentic evolution as an interactive environment . As illustrated in Figure 1 , this view shifts evolution from an unstructured iterative process into an environment that exposes process-level state and supports external intervention. The state is the accumulated evolution context, including candidates, feedback, traces, failures, costs, and search history. The transition mechanism is the current evolution mechanism: either an explicit search procedure or the operating context that shapes a general-purpose agent’s future decisions. A meta-agent acts on this environment not by generating the next candidate, but by editing the mechanism that controls how future evolution proceeds. This makes the same environment view applicable to both hand-designed procedures and general-purpose evolution agents.

Realizing this view requires a harnessed design. The evolution environment is large, noisy, and constantly changing. Without a stable interface, a meta-agent may lose track of reliable evidence, revisit old attempts or make edits whose effects are hard to verify. At the same time, evaluation and candidate records must remain protected from the agents that modify the evolution process. These challenges motivate a harness that makes evolution observable, editable, and externally governed.

We therefore introduce AEvo , a harnessed framework for meta-editing agentic evolution. AEvo standardizes the evolution workspace, protects the evaluator, records every evaluated candidate into a searchable history, and exposes process-level information to the meta-agent. It then runs evolution through a two-phase loop. In the meta-editing phase, the meta-agent edits the current mechanism and specifies how the next segment should run. In the evolution segment, the updated mechanism runs under this plan and produces multiple candidates before the next meta-agent intervention. The same loop can revise both procedure- and agent-based evolution, reducing the risk of local optima.

Our contributions are threefold. (1) Environment Formulation: We formulate agentic evolution as an interactive environment, where accumulated evolution context becomes process-level state and meta-actions edit the mechanism that drives future evolution. (2) Harnessed Meta-Editing: We introduce AEvo , a harnessed framework for meta-editing agentic evolution that protects evaluation, records evaluated candidates, and supports coarse-grained intervention through meta-editing phases and evolution segments. (3) Cross-Form Instantiation and Evaluation: We instantiate the same framework on both procedure-based and agent-based evolution, showing that AEvo can revise either explicit procedure components or agent operating contexts. On standard agentic and reasoning benchmarks, AEvo outperforms five evolution baselines and achieves a 26% relative improvement over the strongest baseline. On three open-ended optimization tasks, AEvo outperforms four evolution baselines and achieves state-of-the-art performance under the same iteration budget.

## 2 Related Work

Agentic Evolution. A growing line of work uses LLMs and agents to iteratively improve artifacts through generation, feedback, and revision [ 15 , 14 , 13 , 28 ] . Prompt methods optimize language-model programs or feedback-driven prompts, including DSPy [ 12 ] , SPO [ 30 ] , TextGrad [ 37 ] , and GEPA [ 1 ] . Another line studies the automated design and evolution of agentic systems and workflows, such as ADAS [ 9 ] , Darwin Gödel Machine [ 40 ] , Huxley-Gödel Machine [ 27 ] , AFlow [ 42 ] , RobustFlow [ 32 ] , and SkillRL [ 29 ] . Recent open-ended discovery systems further apply evolutionary search to scientific and algorithmic discovery, including AlphaEvolve [ 18 ] , OpenEvolve [ 24 ] , TTS-Discover [ 38 ] , CORAL [ 22 ] , SimpleTES [ 36 ] , and ASI-Evolve [ 33 ] . However, their search behavior is typically controlled either by fixed procedures or by agents directly managing candidate generation. In contrast, AEvo treats the evolution process itself as an interactive environment and studies how to steer the mechanism that controls future search.

Agentic Meta-Evolution. Early meta-learning work showed that the learning rule itself can be optimized, for example by learning recurrent reinforcement-learning dynamics [ 26 ] or evolving policy-gradient objectives [ 8 ] . Recent agentic systems extend this idea to editable agent programs and memory systems. HyperAgents study self-referential agent programs in which both task-solving behavior and the meta-improvement mechanism can be modified [ 41 ] . MemEvolve and ALMA similarly explore meta-evolution over agent memory designs [ 39 , 31 ] . Unlike HyperAgents, which internalize meta-improvement within a self-modifying agent program, AEvo treats agentic evolution as an interactive environment observed and edited through an external harness, covering both hand-designed procedures and general-purpose agents while keeping evaluation and candidate recording externally governed.

## 3 Problem Formulation

### 3.1 Agentic Evolution

We formulate agentic evolution as a process for optimizing an artifact through repeated improvement rounds. Let x ∈ 𝒳 x\in\mathcal{X} denote the object being optimized, such as a program, prompt, workflow, skill, tool, or agent component. We use r r to index the evolution round. Each round produces a round context c r c_{r} , which contains the candidates generated in that round, their evaluation results, execution traces, failures, costs, and any intermediate information produced during optimization. The accumulated evolution context after r r rounds is denoted as 𝒞 r = ( c 1 , c 2 , … , c r ) . \mathcal{C}_{r}=(c_{1},c_{2},\ldots,c_{r}). Finally, let Π \Pi denote the optimization mechanism that advances evolution: c r = Π ⁡ ( 𝒞 r − 1 ) , 𝒞 r = 𝒞 r − 1 ⊕ c r , c_{r}=\Pi(\mathcal{C}_{r-1}),\qquad\mathcal{C}_{r}=\mathcal{C}_{r-1}\oplus c_{r}, where ⊕ \oplus appends the newly produced round context to the accumulated evolution context. Π \Pi does not have to be a fixed algorithm; it can also be an agentic process that reads the history, reasons over feedback, and decides how to generate the next candidate. Thus, Π \Pi represents the mechanism by which search is continued from the current evolution context.

Under this formulation, existing agentic evolution methods mainly differ in how Π \Pi is instantiated. In procedure-based evolution , Π \Pi is a predefined outer loop whose behavior is mainly determined by selection and optimization: the selection rule chooses previous candidates or contexts from 𝒞 r − 1 \mathcal{C}_{r-1} , while the optimization operator generates new candidates from the selected information. Evaluation assigns scores, traces, and feedback to the generated candidates, providing signals for future selection and update. In agent-based evolution , Π \Pi is instead implemented by a general-purpose agent. Rather than following fixed selection-and-optimization rules, the agent reads the accumulated context 𝒞 r − 1 \mathcal{C}_{r-1} and decides what to do next, such as inspecting feedback, comparing candidates, modifying artifacts, writing tools, or generating new attempts. Thus, procedure-based evolution specifies search control explicitly but rigidly, while agent-based evolution leaves search control implicit in the agent’s context-conditioned behavior.

In both cases, evolution proceeds by repeatedly applying Π \Pi while accumulating context 𝒞 r \mathcal{C}_{r} . This context records not only the candidates produced by evolution, but also how search has unfolded through evaluation results, traces, failures, costs, and intermediate artifacts. The next subsection uses this accumulated context to define an environment view of evolution.

### 3.2 Evolution as an Interactive Environment

We treat the evolution process itself as an interactive environment for a meta-agent. At round r r , the state of this environment is defined by the round index and the accumulated evolution context: s r = ( r , 𝒞 r ) . s_{r}=(r,\mathcal{C}_{r}). When the optimization mechanism may change across rounds, we write the current mechanism as Π r \Pi_{r} . This mechanism specifies the transition of the environment. Without intervention, the next round is produced by applying the current mechanism to the current context: c r + 1 = Π r ​ ( 𝒞 r ) , s r + 1 = ( r + 1 , 𝒞 r ⊕ c r + 1 ) . c_{r+1}=\Pi_{r}(\mathcal{C}_{r}),\qquad s_{r+1}=(r+1,\mathcal{C}_{r}\oplus c_{r+1}). Thus, Π r \Pi_{r} is the transition rule that determines how the evolution process continues.

To interact with this environment, we introduce a meta-agent M M . The role of Π r \Pi_{r} is to continue the candidate search, while the role of M M is to act on the evolution process that governs this search. Since the full state s r s_{r} can be large and noisy, the meta-agent receives an observation extracted from the state: o r = Φ ⁡ ( s r ) = Φ ⁡ ( r , 𝒞 r ) , o_{r}=\Phi(s_{r})=\Phi(r,\mathcal{C}_{r}), where Φ \Phi summarizes relevant information from the accumulated context, such as progress, repeated failures, invalid attempts, cost patterns, or redundant search directions.

Given this observation, the meta-agent produces an edit action: a r = M ⁡ ( o r ) . a_{r}=M(o_{r}). The action does not directly become the next candidate. Instead, it modifies the transition rule of the evolution environment: Π r + 1 = Edit ⁡ ( Π r , a r ) . \Pi_{r+1}=\mathrm{Edit}(\Pi_{r},a_{r}). The edited mechanism is then used to continue evolution: c r + 1 = Π r + 1 ​ ( 𝒞 r ) , 𝒞 r + 1 = 𝒞 r ⊕ c r + 1 . c_{r+1}=\Pi_{r+1}(\mathcal{C}_{r}),\qquad\mathcal{C}_{r+1}=\mathcal{C}_{r}\oplus c_{r+1}. In this sense, we formulate agentic evolution as an environment in which the state is the accumulated evolution context, the observation is a summary of this context, and the action edits the mechanism that controls future search.

This formulation applies to both forms of agentic evolution. For procedure-based evolution , editing Π r \Pi_{r} changes explicit components such as selection, optimization, feedback use, budget allocation, or update rules. For agent-based evolution, editing Π r \Pi_{r} changes the agentic context that shapes future decisions, such as skills, goals, tools, feedback format, or execution context. In both cases, the meta-agent steers evolution not by proposing one more candidate, but by modifying how subsequent search is carried out. Section 4 describes the system design used to instantiate this formulation.

## 4 Methodology

Figure 2 illustrates AEvo . AEvo instantiates the environment view in Section 3 as a harnessed loop that alternates between a meta-editing phase and an evolution segment . In the meta-editing phase, the meta-agent updates the current evolution mechanism Π r \Pi_{r} and specifies how the next segment should run, including its iteration budget and stopping conditions. In the evolution segment, the updated mechanism runs under this plan and may produce multiple evaluated candidates before the next meta-agent intervention. Thus, one meta-edit can govern a segment of future evolution rather than a single candidate.

### 4.1 Design of AEvo

Meta-editing phase. The meta-editing phase decides both what to change and how to continue . The meta-agent can be any coding-capable agent that can inspect a workspace, edit files, execute commands, and follow the AEvo meta-agent skill specification, such as Claude Code [ 3 ] , Codex [ 19 ] , or open-source coding agents [ 34 , 20 ] . Given the current workspace, it inspects the accumulated history, then produces a meta-action consisting of a workspace edit and a run plan. The workspace edit modifies files that define Π r \Pi_{r} , such as procedure code, prompts, skills, goals, tools, feedback formats, validators, notes, or execution context. The run plan specifies how the next evolution segment should proceed, including the allowed iterations, budget use, and stopping conditions.