Huxley-Gödel Machine: Human-Level Coding Agent Development by an Approximation of the Optimal Self-Improving Machine

Abstract: Recent studies operationalize self-improvement through coding agents that edit their own codebases. They grow a tree of self-modifications through expansion strategies that favor higher software engineering benchmark performance, assuming that this implies more promising subsequent self-modifications. However, we identify a mismatch between the agent's self-improvement potential (metaproductivity) and its coding benchmark performance, namely the Metaproductivity-Performance Mismatch. Inspired by Huxley's concept of clade, we propose a metric ($\mathrm{CMP}$) that aggregates the benchmark performances of the descendants of an agent as an indicator of its potential for self-improvement. We show that, in our self-improving coding agent development setting, access to the true $\mathrm{CMP}$ is sufficient to simulate how the G\"odel Machine would behave under certain assumptions. We introduce the Huxley-G\"odel Machine (HGM), which, by estimating $\mathrm{CMP}$ and using it as guidance, searches the tree of self-modifications. On SWE-bench Verified and Polyglot, HGM outperforms prior self-improving coding agent development methods while using less wall-clock time. Last but not least, HGM demonstrates strong transfer to other coding datasets and LLMs. The agent optimized by HGM on SWE-bench Verified with GPT-5-mini and evaluated on SWE-bench Lite with GPT-5 achieves human-level performance, matching the best officially checked results of human-engineered coding agents. Our code is available at https://github.com/metauto-ai/HGM.

• The paper introduces clade-based metaproductivity, a metric that reliably guides long-term self-improvement in coding agents.
• The paper models self-improvement as an iterative tree search, decoupling expansion from evaluation to optimize resource use.
• The paper demonstrates superior performance on benchmarks, achieving higher accuracy with significantly reduced CPU-hours.

Huxley-Gödel Machine: Clade-Based Self-Improvement for Human-Level Coding Agents

Introduction and Motivation

The Huxley-Gödel Machine (HGM) introduces a new paradigm for the development of self-improving coding agents by addressing a critical limitation in prior approaches: the misalignment between immediate benchmark performance and long-term self-improvement potential. Existing methods, such as Darwin Gödel Machine (DGM) and Self-Improving Coding Agent (SICA), select self-modifications based on short-term benchmark gains, implicitly assuming that higher immediate performance correlates with greater future improvement. However, empirical evidence demonstrates a weak correlation between these metrics and the actual productivity of agent lineages, a phenomenon termed the Metaproductivity–Performance Mismatch (MPM).

To resolve this, HGM proposes a lineage-based metric, Clade-Metaproductivity (CMP), inspired by Huxley’s concept of clades in evolutionary biology. CMP aggregates the downstream success of all descendants of an agent, providing a more reliable estimate of its long-term self-improvement capacity. Theoretical analysis shows that, under certain assumptions, access to a true CMP oracle suffices to simulate the optimal acceptance mechanism of the Gödel Machine in the context of coding agent development.

Formalization of Self-Improvement as Tree Search

HGM models the self-improvement process as an iterative tree search, where each node represents an agent and edges correspond to self-modifications. The search policy alternates between expanding the tree (generating new agents via self-modification) and evaluating existing agents on downstream tasks. This compound policy is decomposed into three sub-policies: selection (expansion vs. evaluation), expansion (which agent to modify), and evaluation (which agent to test).

Unlike DGM and SICA, which tightly couple expansion and evaluation, HGM decouples these steps, enabling asynchronous and fine-grained control. This allows for early stopping on unpromising agents and more efficient allocation of computational resources.

Clade-Metaproductivity and Theoretical Foundations

The core theoretical contribution is the definition of CMP as a localized variant of global metaproductivity (GMP). While GMP measures the expected utility of the final agent across the entire tree, CMP focuses on the subtree (clade) rooted at a given agent. Formally, for a policy $\pi$ and agent $a$ in tree $\mathcal{T}$:

$$\mathrm{CMP}_\pi(\mathcal{T}, a) = \mathbb{E}_{\mathcal{T}_B \sim p_\pi(\cdot \mid \mathcal{T}, a)} \left[ \max_{a' \in C(\mathcal{T}_B, a)} U(a') \right]$$

where $C(\mathcal{T}_B, a)$ denotes the clade rooted at $a$ in the final tree $\mathcal{T}_B$, and $U$ is the utility function (e.g., average task success).

A key result is that, under the assumptions of repeatable trials, fixed evaluation environments, and budgeted self-modification, access to a CMP oracle is sufficient to implement the Gödel Machine’s optimal acceptance mechanism. This establishes CMP as a theoretically justified guidance metric for self-improvement in coding agents.

HGM Algorithmic Framework

HGM operationalizes these insights via a structured policy:

• Expansion Policy: Agents are selected for expansion based on Thompson sampling over their estimated CMP, computed as the weighted average of successes and failures across their clade. This probabilistic approach balances exploration and exploitation, with an adaptive scheduler increasing exploitation as the budget is consumed.
• Evaluation Policy: Agents are selected for evaluation using Thompson sampling over their individual empirical performance, prioritizing those with higher estimated utility.
• Selection Policy: The decision to expand or evaluate is governed by a UCB-Air-inspired rule, which triggers expansion when the number of evaluations grows superlinearly with the number of agents.

The decoupling of expansion and evaluation enables asynchronous execution, allowing HGM to fully utilize available computational resources and reduce wall-clock time.

Figure 1: Visualization of the pull force in the self-improvement tree, illustrating the dynamics of agent expansion and evaluation.

Empirical Analysis: Metaproductivity–Performance Mismatch

Empirical studies on SWE-bench Verified and Polyglot benchmarks reveal a weak correlation between immediate benchmark performance (as used by DGM and SICA) and empirical CMP, confirming the MPM hypothesis. In contrast, HGM’s CMP estimator achieves substantially higher correlation with true metaproductivity, both in weighted and unweighted settings.

Figure 2: (Left) Weak correlation between benchmark-based guidance and long-term self-improvement; HGM’s clade-level metric mitigates this mismatch. (Right) HGM achieves higher accuracy with 2.38× less CPU time on SWE-bench Verified.

Self-Improvement and Efficiency

HGM demonstrates superior self-improvement capability compared to DGM and SICA. On SWE-bench Verified-60 and Polyglot, HGM achieves the highest final agent accuracy (56.7% and 30.5%, respectively), with significant efficiency gains—requiring 2.38× and 6.86× fewer CPU-hours than DGM on the respective benchmarks. SICA is further hampered by repeated errors, highlighting the robustness of HGM’s asynchronous, decoupled approach.

Human-Level Coding Agent Design and Generalization

HGM’s best-belief agent, optimized on SWE-bench Verified with GPT-5-mini, matches or surpasses the best human-engineered coding agents on SWE-bench Lite when evaluated with GPT-5. Notably, the agent generalizes robustly to both dataset and model shifts, maintaining high performance on unseen tasks and with larger LLM backbones. This transferability indicates that HGM’s evolutionary process discovers genuinely generalizable agent designs, rather than overfitting to specific benchmarks or models.

Practical and Theoretical Implications

The introduction of clade-based metaproductivity as a guidance metric for self-improvement has several implications:

• Algorithmic Design: CMP provides a theoretically grounded alternative to greedy benchmark-based heuristics, enabling the discovery of agents with greater long-term potential.
• Resource Efficiency: The asynchronous, decoupled policy structure of HGM allows for more efficient use of computational resources, which is critical for large-scale agent evolution.
• Generalization: The lineage-based approach fosters the emergence of agent designs that are robust to distributional and architectural shifts, a key requirement for practical deployment.
• Theoretical Foundations: The formal equivalence between CMP-guided search and the Gödel Machine’s optimal acceptance mechanism (under specified assumptions) bridges the gap between theoretical optimality and practical implementability in self-improving systems.

Future Directions

The clade-based perspective on self-improvement opens several avenues for further research:

• Relaxing Assumptions: Extending the theoretical framework to settings with non-repeatable trials, non-stationary environments, or partial observability.
• Hierarchical and Multi-Objective Optimization: Incorporating additional objectives (e.g., interpretability, safety) into the CMP framework.
• Open-Endedness and Quality Diversity: Integrating quality-diversity optimization to promote the discovery of diverse, high-potential agent lineages.
• Meta-Learning and Transfer: Leveraging CMP-guided evolution for meta-learning update rules or transfer across domains beyond software engineering.

Conclusion

The Huxley-Gödel Machine establishes a lineage-based, theoretically justified approach to self-improving coding agent development. By leveraging clade-level metaproductivity as a guidance metric, HGM overcomes the limitations of benchmark-driven heuristics, achieving both higher agent quality and greater computational efficiency. The demonstrated generalization to new datasets and LLM backbones underscores the potential of clade-based self-improvement as a foundation for scalable, robust, and autonomous agent design. This work suggests that future progress in agentic AI will benefit from a shift toward metrics and algorithms that explicitly account for the long-term generative potential of entire agent lineages, rather than focusing solely on immediate performance gains.