acceptodds
Under review as a conference paper at ICLR 2027

AgentGA: Evolving Code Solutions in Agent-Seed Space

Abstract

We present AgentGA, a framework that evolves autonomous code-generation runs by optimizing the agent seed: the task prompt plus optional parent archives that initialize a fresh workspace. The outer loop searches over reusable starting conditions rather than editing code directly. Each generation launches a fresh autonomous run in an isolated workspace, while selected parent archives provide inherited artifacts that descendants can inspect and reuse. AgentGA couples a population-level genetic algorithm with long-horizon agents, deterministic 1:1 elite tournaments, and online operator allocation through a modified Hedge controller. We evaluate the approach for tabular AutoML on the 16-competition Weco-Kaggle Lite benchmark, including model-matched AIDE and AutoGluon comparisons. AgentGA with DeepSeek V4 Flash (0731) achieves 71.13% mean Exceeds % of Human, compared with 57.61% for model-matched AIDE and 57.89% for AutoGluon, outranking AIDE on 14/16 tasks. These descriptive results support agent-seed optimization as a practical design choice for autonomous code-search systems, without isolating the causal contribution of inheritance.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
LLM agents, autonomous agents, evolutionary algorithms, genetic algorithms, AutoML, code generation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_4jzrOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents AGENT GA, a framework that uses genetic algorithms to improve solutions generated by autonomous coding agents. Instead of directly evolving code, the framework evolves the initial conditions of agent runs, including task prompts and archived solutions from previous generations. Multiple agents independently explore and improve solutions, while an evolutionary algorithm selects better results and adjusts the probabilities of different search operations using a modified HEDGE algorithm. Experiments on 16 Kaggle competitions show that AGENT GA achieves higher average leaderboard rankings than AIDE, AutoGluon, and a single-agent baseline.

Critical strengths

The paper addresses a valuable problem. Improving the effectiveness of autonomous coding agents on long-running optimization tasks is important. Allowing agents to reuse previous solutions while exploring different approaches is a practical direction.

Critical weaknesses
  1. The methodological contribution is not sufficiently established. The concern is that the paper does not clearly show what specific challenges of autonomous agent search require the proposed design, or how its adaptations address these challenges beyond existing approaches. The paper identifies early termination and the difficulty of maintaining useful information across long agent runs. It addresses these issues by restarting agents, passing previous solutions through files, and applying population-based search. These are reasonable design choices, but the paper does not explain why they require the proposed evolutionary framework rather than a simpler restart-and-reuse approach. Moreover, the paper defines agent-seed optimization over stochastic agent executions, but the algorithm does not explicitly address the uncertainty in evaluating different seeds. As a result, the main methodological insight of the proposed framework remains unclear.

  2. The experimental comparisons are not fair in terms of computational resources. Although AGENT GA significantly outperforms AIDE, the two methods use very different computational budgets. As acknowledged in Appendix G.1, AIDE uses approximately three orders of magnitude fewer LLM tokens. Therefore, the improvement could simply result from substantially more computation and search attempts rather than the proposed framework. The single-agent comparison has a similar problem. Although both methods are given the same nominal attempt allowance, most single agents stop early. AGENT GA actually performs 4.7 times as many scored experiments and incurs 20.7 times the LLM cost. This comparison does not establish that the proposed evolutionary framework is more effective under a comparable budget.

  3. The experiments do not establish the contribution of the proposed components. The paper does not provide sufficient ablation studies to show why the proposed framework works. In particular, it does not isolate the effects of population-based search, archive inheritance, genetic operators, or HEDGE-based adaptive allocation. The reported 70.7% win rate of agents inheriting previous solutions, compared with 17.8% for agents starting from scratch, is also not convincing evidence for the proposed inheritance mechanism. Agents with inherited solutions start from previously optimized results, while the others must start over. This difference is expected and does not demonstrate the value of the specific design. More importantly, the paper does not compare AGENT GA against a simple approach that repeatedly restarts an agent with the best previous solution. Such a baseline would help determine whether the proposed genetic algorithm is necessary at all.

Questions for authors
  1. What specific challenge of autonomous agent search does the proposed evolutionary framework solve that cannot be addressed by simpler approaches? In particular, why are population-based evolution and the proposed agent-seed formulation necessary beyond repeatedly restarting agents with previous solutions?

  2. How does AGENT GA compare with AIDE and other baselines under the same computational budget? Can the authors provide comparisons with matched LLM token costs or numbers of scored experiments?

Reviewer_bw7SOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes AGENTGA, a genetic algorithm framework for autonomous code generation. Instead of directly evolving code artifacts, it evolves the initial conditions of autonomous agent runs, consisting of task instructions and inherited archives from previous solutions. Each generation launches fresh agents that independently explore solutions, while a population-based evolutionary process selects successful descendants and adaptively allocates different search operators.

Critical strengths

What I like about the approach is that it is a practical approach to sustained agentic search, i.e., resetting conversations while transferring useful artifacts through archives is an interesting solution to context growth and premature agent termination. Furthermore, I find the experiments showing that a single agent frequently stops after only a small fraction of its allocated attempts are particularly interesting.

Critical weaknesses
  • For me the distinction between evolving agent seeds, code artifacts, and agent implementations is reasonable. However, population-based evolution of autonomous code-search agents with inherited solutions and experience has already been explored: GEAR (Genetic AutoResearch, 2026) and EvE (Evolutionary Ensemble of Agents, 2026). I might have missed something but I think the authors need to establish what substantive algorithmic contribution AGENTGA introduces beyond existing evolutionary approaches, rather than relying primarily on the distinction in the representation of the search space.

  • The comparison with AIDE is not compute-matched. Moreover, from my understanding of section Section 4.4, although the single-agent control is given the same nominal experiment allowance, AGENTGA executes 4.7× more scored experiments at 20.7× the LLM cost. Consequently, the comparison does not isolate the benefits of population-based evolution from additional computation or repeated agent restarts.

Questions for authors

1- How does AGENTGA fundamentally differ from GEAR and EvE, which also employ population-based evolutionary search with autonomous agents and inherited research experience? What capability or algorithmic advantage is uniquely enabled by optimizing agent seeds?

2- Could GEAR or a comparable population-based autonomous code-search approach be adapted to the same AutoML setting? If not, what prevents a meaningful comparison?

3- How does AGENTGA compare with AIDE when both are given comparable LLM and evaluation budgets? To what extent can the improvements be explained by additional search?

4- How would AGENTGA compare with a simpler iterative agent that is repeatedly restarted, inherits its own previous solutions and experiments, and receives the same compute budget? This would help determine whether maintaining multiple evolving lineages is necessary.

Reviewer_nsMkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AGENTGA, a framework that evolves agent seeds rather than directly modifying code solutions. It combines population-based genetic search, inherited parent archives, elite tournaments, and adaptive operator allocation. Experiments on 16 Weco-Kaggle Lite competitions show that AGENTGA outperforms AIDE and AutoGluon, achieving 71.13% mean Exceeds % of Human and surpassing AIDE on 14 of 16 tasks.

Critical strengths

The paper explores a useful direction by optimizing the initial conditions of autonomous agent runs rather than directly editing code solutions. The combination of fresh workspaces and inherited archives provides an interesting perspective on evolutionary agent-based code generation.

Critical weaknesses
  1. The evaluation is restricted to a single domain (16 tabular Kaggle competitions), one LLM backbone (DeepSeek V4 Flash), and one custom agent harness. Moreover, these are older public competitions with potential pretraining contamination, as acknowledged in Appendix G.3. The paper does not evaluate on broader benchmarks.

  2. The proposed agent-seed formulation is closely related to existing archive-based evolutionary search, where agents iteratively improve solutions using previous artifacts and feedback. Moreover, the population of three with deterministic 1:1 elite tournaments resembles parallel elitist search with occasional cross-lineage recombination. The paper does not establish the necessity of these evolutionary mechanisms through controlled ablations. Closely related methods, including AIRA-dojo, ML-Master, and R&D-Agent, are neither discussed nor compared. In particular, AIRA-dojo has already explored greedy, MCTS, and evolutionary search with agentic operators on MLE-bench. The authors should better clarify the conceptual differences and demonstrate the unique benefits of the proposed formulation.

Questions for authors

AIDE (DS) was run at its default 20-step budget and used roughly three orders of magnitude fewer tokens than AgentGA, while AgentGA 1×1 used 20.7× less LLM cost. Could you add compute-matched baselines to separate the contribution of inheritance and population structure from that of extra compute? Two would be especially informative:

An inheritance-free restart baseline: best-of-N independent Initial runs with a total budget matched to AgentGA.

AIDE run with an increased step budget, so that its total tokens are comparable to AgentGA.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.