Harnessing Agentic Evolution — part 2/3

Cleaned-text code points [15937, 31637). Bibliography omitted here; source.md includes it.

This design makes the meta-agent a process-level editor rather than a candidate generator: it changes the mechanism and conditions under which future candidates are produced. When evolution is productive, the meta-agent may allocate more iterations to the current mechanism; when it repeatedly produces invalid candidates, redundant attempts, or irrelevant exploration, the meta-agent may stop the segment and revise Π r \Pi_{r} before continuing. The full meta-agent skill specification and pseudocode of the meta-editing loop are provided in Appendix C.2.3 .

Harnessed evolution segment. An evolution segment is the interval executed after a meta-edit. It runs the current mechanism Π r \Pi_{r} under the run plan produced by the meta-agent. Depending on the setting, this segment may consist of several rounds of a procedure, or an inner-agent session that produces multiple candidate attempts. Each candidate submitted for official evaluation passes through the harness-controlled evaluator, and the resulting artifact, score, trace, failure information, cost, and provenance are appended to the candidate history.

The harness provides the stable boundary needed for reliable meta-editing. It organizes candidates, logs, traces, evaluation records, meta-agent instructions, and editable evolution components into a fixed workspace layout. To prevent reward hacking, the evaluator is isolated from both the evolution agent and the meta-agent: agents can submit candidates, but they cannot inspect evaluator internals, access hidden benchmark artifacts, or directly write official scores. The harness further exposes a command-line interface for initializing workspaces, launching evolution segments, inspecting recent status and candidate history, and continuing the current process. Thus, the harness does not decide how evolution should improve; it provides the protected and inspectable interface through which evolution can be observed, edited by the meta-agent, recorded, and resumed.

### 4.2 Instantiating AEvo

AEvo applies the same two-phase loop to both forms of agentic evolution. The outer loop is unchanged: the meta-agent edits the current mechanism Π r \Pi_{r} and specifies how the next evolution segment should run. The difference lies in what Π r \Pi_{r} consists of.

Procedure-based evolution. For procedure-based evolution, Π r \Pi_{r} is an explicit evolution procedure. It defines how previous candidates or contexts are selected, how new candidates are generated, how evaluation feedback is used, and how candidate history is updated. A meta-action therefore edits the procedure itself, such as revising the selection strategy, changing the optimization operator, altering the feedback summary, adding local filtering or retry logic, adjusting budget use, or repairing candidate management. The edited procedure then controls the next evolution segment, which may run for multiple rounds before the next meta-agent intervention.

Agent-based evolution. For agent-based evolution, Π r \Pi_{r} is the operating context of a general-purpose evolution agent, including goals, skills, tools, memory files, shared notes, validators, and execution setup. A meta-action therefore edits the conditions under which the next inner-agent session will evolve, such as revising a skill, rewriting the session goal, changing how evaluator feedback is presented, or reorganizing shared notes. The inner agent remains responsible for generating candidates, while AEvo revises the context that shapes future evolution.

## 5 Experiments

### 5.1 Settings

Tasks. We evaluate AEvo on two standard benchmarks, Terminal-Bench [ 25 ] and ARC-AGI-2 [ 5 ] , and three open-ended optimization tasks, circle_packing_26 (CP26) [ 21 ] , autocorrelation_second (AC2) [ 4 ] , and Anthropic’s Kernel optimization task [ 2 ] . Together, these tasks cover agentic problem solving, abstract reasoning, and code-evolving open-ended optimization. Detailed task definitions and evaluation protocols are given in Appendix C.1 .

Baselines. We group baselines into three families: single-agent inference, agent-based evolution, and procedure-based evolution. On Terminal-Bench and ARC-AGI-2, we compare against one-shot ReAct [ 35 ] and five procedure-based evolution baselines: ADAS [ 9 ] , DGM [ 40 ] , AFlow [ 42 ] , SPO [ 30 ] , and GEPA [ 1 ] . On the three open-ended tasks, we compare against two agent-based evolution baselines, Codex and Claude Code, and two procedure-based evolution baselines, OpenEvolve [ 24 ] and HyperAgents [ 41 ] . This setup lets us compare both variants of AEvo against systems that either keep a fixed search procedure or rely on an agent to improve artifacts directly.

Implementation Details. AEvo is instantiated in two forms. In the procedure-based setting, a meta-agent edits the evolution procedure while leaving the task evaluator fixed. In the agent-based setting, the meta-agent steers a coding-agent harness through prompts, notes, and reusable utilities. We use Claude Code and Codex as the meta-agent interfaces, backed by Claude-Opus-4.7 and GPT-5.4 as the optimization models. For Terminal-Bench and ARC-AGI-2, candidate execution uses Gemini-3-Flash. All models are accessed through APIs. Full hyperparameter settings, round budgets, early-stopping criteria, and initialization details are given in Appendix B .

Metrics. For Terminal-Bench and ARC-AGI-2, we report task score and the first optimization round that reaches the best score. Results are summarized with Avg@3 over three independent runs. For the open-ended tasks, we report the task-native objective with Best@3 over three runs, together with the first round that reaches the best result (Best R.) and the average dollar cost per optimization round ($/R). Exact task-specific objectives and cost computation are provided in Appendix C.1 .

### 5.2 Main Results

Overall performance. Tables 1 and 2 show that AEvo consistently improves agentic evolution across both open-ended optimization and fixed benchmarks. On the open-ended tasks, AEvo achieves the best or tied-best result on all three tasks, while also improving the speed or stability with which strong candidates are found. In particular, on the Kernel optimization task, AEvo achieves 1138 cycles within 100 iterations, which is, to our knowledge, the best reported result under the same iteration budget. This suggests that the harnessed meta-editing loop improves how evolution uses feedback over time, rather than merely increasing the number of candidate attempts. On standard agentic and reasoning benchmarks, AEvo also improves procedure-based evolution over strong fixed-loop baselines. The gains are consistent across both Terminal-Bench and ARC-AGI-2, yielding a 26% relative improvement over the strongest baseline on average. Together, these results support the central claim that mechanism-level intervention can benefit both open-ended optimization and benchmark-driven agentic evolution.

Improvement through optimization-time reasoning. The gains on Table 2 come with a higher per-round optimization cost. AEvo costs about three times as much as procedure-based baselines on these benchmarks. This means that AEvo improves performance by scaling the reasoning and deliberation used during optimization, and supports our view that increasing the budget of the evolution process is a useful axis for improving agentic evolution.

Cost analysis and agentic behavior. The open-ended tasks reveal that cost is not solely determined by whether a method is agent-based or procedure-based. Agent-based evolution can remain cost-competitive when implemented through coding-agent interfaces with prompt caching and persistent contexts. AEvo Agent {}_{\texttt{Agent}} maintains low per-round cost: 0.34–0.32 on circle_packing_26 , 1.40–1.31 on autocorrelation_second , and 1.27–1.23 on Kernel optimization. By contrast, procedure-based methods can become expensive in long-horizon optimization by repeatedly constructing large prompts over an expanding search history without comparable caching, as visible in HyperAgents’ higher per-round cost on Task 1 and Task 2.

At the same time, direct coding agents show why agent freedom alone is insufficient for reliable evolution. Even with prompts encouraging long-horizon search, coding agents often stop early once local improvements become difficult, as seen in the early best rounds of Codex on Task 1 and Task 3. This suggests that an agent’s internal stopping decision can conflict with the external evolution budget. AEvo addresses this by placing the coding agent inside an explicit evolution harness, where rounds, candidate records, and evaluation feedback are maintained outside the agent’s local decision loop.

### 5.3 Evolution Dynamics

Revising evolution after plateaus. As shown in Figure 3 , procedure-based methods such as OpenEvolve and HyperAgents tend to flatten once their current selection or mutation strategy stops producing useful candidates. In contrast, AEvo can revise the mechanism that drives subsequent evolution. When progress stalls or repeated failures appear, these signals become process-level feedback: the meta-agent can adjust the procedure, directive, or reusable search context, producing step-wise improvements after plateaus. This is visible in the late-stage jump that leads to the best 100-round result.

Using the evolution budget effectively. Direct coding agents can obtain strong early gains through internal simulation, execution, and debugging, but they may stop early once local progress becomes difficult. AEvo avoids tying progress to the agent’s local stopping decision by maintaining explicit rounds, candidate records, and evaluation feedback outside the agent context. This allows the external evolution budget to be used more consistently.

Scaling beyond early gains. The right panel extends Codex-based AEvo from 100 to 200 iterations. The best result improves from 1138 to 1121 cycles, showing that AEvo continues to benefit from additional rounds rather than saturating after an early strong candidate. Overall, the trajectories suggest that agent flexibility is useful, but reliable long-horizon improvement requires a harness that preserves global evidence and enables mechanism-level correction.

### 5.4 Ablation Study

Table 3 (Appendix) ablates two key components of AEvo Agent {}_{\texttt{Agent}} on the Kernel optimization task. The full system completes the 100-round budget without reward hacking and reaches the best valid result of 1138 cycles. Removing meta-agent skills does not lead to reward hacking, but it substantially weakens long-horizon search: the best run only reaches 1407 cycles, and the runs do not consistently sustain the full budget. Removing the evolution harness is even less reliable. Although one run finds a strong 1167-cycle solution, two of the three runs enter reward-hacking trajectories and fail to produce valid cycle results. These results suggest that the skills mainly support sustained and effective meta-intervention, while the harness provides the protected evaluation boundary and structured evolution context needed to keep agentic search aligned with the true objective.

### 5.5 Case Study

Figure 4 illustrates how AEvo performs meta-intervention during procedure-based evolution on an ARC-AGI-2 task. Starting from P 0 P_{0} , the meta-agent initializes a best-parent rewrite procedure that selects candidate agents by validation accuracy. This first produces an initial breakthrough candidate C 1 C_{1} , but subsequent variants expose several failure modes in observation parsing and refinement. The meta-agent then revises the procedure rather than continuing the same search blindly: P 1 P_{1} adds Pass@K sampling and local scoring for verifier-guided generation, P 2 P_{2} fixes observation parsing to activate feedback-guided refinement, and P 3 P_{3} extends the refinement horizon to use more pass/fail feedback before submission. When the search becomes stuck, P 4 P_{4} drops stale feedback and samples more diverse alternatives, leading to a stronger candidate. Later interventions P 5 P_{5} – P 6 P_{6} explore stronger de-anchoring through task-profile and skeleton prompts, but these regress from P 4 P_{4} . This example shows that failed candidates are not merely discarded; they become process-level evidence that helps the meta-agent decide how to revise the future evolution procedure. Additional evolved procedures, optimized agent harnesses, and task-level analyses are provided in Appendix C.3.1 and Appendix C.3.2 .

## 6 Conclusion

We presented AEvo , a harnessed framework for steering agentic evolution by treating the evolution process itself as an interactive environment. Instead of generating one more candidate, AEvo exposes accumulated candidates, feedback, traces, failures, costs, and search history as process-level evidence, and uses a meta-agent to edit the mechanism that controls future evolution. This formulation provides a unified view of procedure-based and agent-based evolution: the same meta-editing loop can revise explicit search procedures or the operating context of general-purpose evolution agents, while keeping evaluation and candidate recording protected by an external harness. Across agentic, reasoning, and open-ended optimization tasks, AEvo improves over strong fixed-procedure and agent-based baselines, suggesting that long-horizon evolution benefits not only from stronger candidate generators, but also from mechanism-level intervention over how search proceeds. Future work should study more diverse evolution environments, cheaper meta-intervention strategies, and safer deployment of harnessed agentic evolution for scientific discovery, software engineering, and autonomous code optimization.

## Appendix A Ablation Study Details

## Appendix B Implementation Details

When temperature is exposed, we set it to 1; when reasoning-effort control is available, we use the high setting; and we use a maximum context budget of 128k tokens. We run 20 optimization rounds on Terminal-Bench and ARC-AGI-2. On open-ended optimization tasks, we allow up to 100 rounds and early-stop when a run reaches the known target score or fails to improve for 25 consecutive rounds. The initial procedure in the procedure-based setting uses best-valid-candidate selection plus a heuristic LLM optimizer; details are given in Appendix C.2.1 . The initial agent used on the standard agentic and reasoning tasks is a ReAct-style [ 35 ] agent; details are given in Appendix C.2.2 .

## Appendix C Additional Experimental Details

### C.1 Task and Metric Details

Our main experiments cover two standard benchmarks and three open-ended optimization tasks. Terminal-Bench evaluates end-to-end task completion in terminal environments, while ARC-AGI-2 measures abstract reasoning under fixed evaluation rules. The three open-ended tasks use hidden or fixed external evaluators and require the optimizer to improve executable code rather than only produce one-shot answers.

##### Circle Packing.

In circle_packing_26 , the goal is to pack 26 circles into a unit square and maximize the sum of their radii. The evaluator returns a validity bit and the achieved packing score. Higher is better.

##### Autocorrelation-Second.

In autocorrelation_second , the goal is to construct a non-negative function on [ − 1 / 4 , 1 / 4 ] [-\nicefrac{{1}}{{4}},\nicefrac{{1}}{{4}}] that maximizes R ⁡ ( f ) = ‖ f ∗ f ‖ 2 2 ‖ f ∗ f ‖ 1 ⋅ ‖ f ∗ f ‖ ∞ . R(f)=\frac{\|f*f\|_{2}^{2}}{\|f*f\|_{1}\cdot\|f*f\|_{\infty}}. The evaluator scores the submitted construction directly by this ratio. Higher is better.

##### Performance Engineering.