The paper introduces HyperAgents and the Darwin Gödel Machine with Hyperagents (DGM-H), a novel framework for open-ended, self-accelerating artificial intelligence that can improve its own learning and problem-solving processes across any computable task. Traditional self-improving AI systems, including the prior Darwin Gödel Machine (DGM), are limited by fixed, handcrafted meta-level mechanisms that constrain how improvements can compound and generalize. While DGM demonstrated recursive self-improvement in coding by having an agent generate and evaluate self-modified variants, its capacity was bottlenecked by a non-modifiable instruction-generation mechanism. Furthermore, DGM's success relied on a limiting assumption: that the skills required for the evaluation task (coding) were intrinsically aligned with those required for self-reflection and self-modification.

HyperAgents address these limitations by integrating a task agent (which solves the target task) and a meta agent (which modifies itself and the task agent) into a single, editable program. This allows the mechanism responsible for generating improvements to be *itself* subject to modification, a process termed metacognitive self-modification. A hyperagent is defined as any computable program, potentially incorporating foundation models (FMs), external tools, or learned components. The task agent is evaluated on a given task (e.g., code modification, paper review decision, robotics reward function design), while the meta agent, given access to an archive of previous agents and evaluation results, proposes changes intended to improve future performance. Crucially, these modifications can target not only the task-solving logic but also the meta agent itself, enabling improvements to the procedures by which future modifications are generated.

The DGM-H instantiates this concept by augmenting the original DGM's open-ended, population-based exploration process with hyperagents. The DGM-H maintains an archive of progressively improving hyperagents, initialized with a single, simple hyperagent built around a frozen FM with external tool-use capabilities. The process alternates between two phases:

1. Metacognitive Self-modification: Parent hyperagents are probabilistically selected from the archive, proportional to their performance and inversely proportional to the number of successful child compilations, biasing towards well-performing hyperagents that generate strong descendants while preserving exploration. The selected parent hyperagent then uses its internal meta agent to generate a modified version of itself. This modification can alter any part of the hyperagent's codebase, including its task-solving logic and its self-improvement mechanism.
2. Evaluation: Each newly generated hyperagent is empirically evaluated on the target tasks. Performance is first estimated on a small subset of tasks, and if sufficient, expanded to a larger set. The evaluated hyperagent is then added back to the archive, serving as a potential stepping stone or parent for future generations.

This iterative process enables the DGM-H to evolve not only how it solves tasks but also *how* it improves itself. By making the entire meta-level mechanism editable, the DGM-H removes the assumption of domain-specific alignment, theoretically allowing self-accelerating progress on any computable task.

The experimental setup validates DGM-H across diverse domains:

• Coding (Polyglot benchmark): Agent modifies code repositories based on natural language instructions.
• Paper Review: Agent predicts binary accept/reject decisions for AI research papers, aligning with human judgments. This tests performance in a hard-to-verify, subjective domain.
• Robotics Reward Design: Agent generates reward functions for robotic tasks, which are then used to train a quadruped robot in simulation. This demonstrates leveraging external simulators and training algorithms.
• Olympiad-level Math Grading: Agent grades math solutions, serving as a testbed for cross-domain transfer of meta-level improvements.

DGM-H is compared against several baselines:

• DGM-H without self-improving meta agents: The meta agent remains fixed to its initial implementation, isolating the contribution of metacognitive self-modification.
• DGM-H without open-ended exploration: No archive is maintained; each new hyperagent replaces its predecessor, isolating the role of population-based search.
• Original DGM: The prior system designed specifically for coding, which has a fixed, handcrafted instruction-generation mechanism.
• DGM-custom: Original DGM manually customized for target domains to provide a stronger comparison, assessing the reliance on human engineering.
• Static Solutions: Handcrafted solutions from prior work for each domain.

The paper reports that DGM-H achieves substantial and generalizable improvements in both task performance and self-improvement ability. On coding tasks, it matches prior state-of-the-art DGM performance despite not being handcrafted for it. Beyond coding, DGM-H significantly outperforms prior self-improving algorithms that struggle outside coding, demonstrating gains in paper review and robotics reward design. Ablations confirm the necessity of both self-improving meta agents and open-ended exploration. Crucially, DGM-H learns transferable meta-level mechanisms (e.g., persistent memory, performance tracking) that systematically improve its ability to generate better task or meta agents over time. These meta-level improvements learned in one domain (paper review/robotics) transfer effectively to another (Olympiad-level math grading) and can compound across runs, suggesting the potential for unbounded, open-ended self-improvement. The authors emphasize safety precautions taken (sandboxing, human oversight) and discuss the broader implications of such self-improving systems.