Hyperagents (DGM-H) is a self-improving agent architecture developed by Meta that extends the Darwin-Gödel Machine framework by unifying task agents and meta-agents into a single editable program. The system enables agents to perform metacognitive self-modification, allowing them to rewrite both their task execution logic and their own evaluation and improvement mechanisms. This architecture represents a significant advancement in agent autonomy and self-directed optimization capabilities.
Hyperagents fundamentally differs from traditional agent systems by eliminating the separation between task agents—which execute primary objectives—and meta-agents—which oversee evaluation and improvement. Instead, both components exist within a unified, editable program structure. This integration allows agents to modify their own code dynamically during execution and learning phases 1).
The system builds upon the Darwin-Gödel Machine framework, which provided foundational concepts for self-referential learning systems. While DGM is primarily optimized for coding tasks and struggles to generalize self-improvement beyond coding, Hyperagents (DGM-H) extends the approach by allowing agents to rewrite their own evaluation logic, enabling metacognitive self-modification that works across diverse domains 2). By consolidating agent hierarchies into a single modifiable program, Hyperagents reduce architectural complexity while increasing the agent's capacity for self-directed improvement. This design enables both task-level optimization and meta-level optimization to occur within the same execution context, facilitating tighter integration between performance and self-modification mechanisms.
Hyperagents has demonstrated substantial performance improvements across diverse domains. In academic paper review tasks, the system improved accuracy from 0.0 to 0.710, indicating successful learning and optimization of evaluation criteria from minimal baselines 3).
In robotics applications, Hyperagents refined a quadruped robot's reward function from an initial performance of 0.060 to 0.372, substantially outperforming human-designed baseline reward functions. This result suggests that the system's capacity for self-modification enables discovery of more effective objective specifications than manually engineered alternatives. The improvement across heterogeneous domains—academic evaluation and robotic control—indicates the architecture's generalizability and adaptability to different problem structures.
The core capability of Hyperagents involves rewriting both task logic and meta-level improvement logic. Unlike systems where improvement mechanisms are fixed, Hyperagents can modify the processes that guide its own optimization. This creates a recursive improvement loop where the agent's evaluative frameworks become subject to the same optimization procedures as its primary task execution.
Self-modification operates on the principle of treating code—whether task logic or improvement logic—as editable entities. Agents can analyze their current performance, identify deficiencies in their own reasoning or evaluation procedures, and implement modifications to address these deficiencies. This metacognitive capacity parallels human meta-learning but at the machine code level, allowing agents to reshape their own computational processes.
Traditional multi-agent systems often employ hierarchical structures where meta-agents operate on fixed evaluation principles while task agents execute within predetermined constraints. Hyperagents eliminates this boundary, creating a unified system where all components remain malleable. This differs substantially from reinforcement learning approaches where the reward function and learning algorithm typically remain static, with only the policy being modified during training 4).
The architecture also contrasts with instruction tuning and supervised fine-tuning approaches, where modifications occur through external feedback rather than internal self-directed mechanisms. By enabling agents to modify their own code directly, Hyperagents achieves a more autonomous form of continuous improvement.
Current applications span both symbolic and embodied domains. Academic paper review represents a symbolic reasoning task requiring subjective evaluation criteria that the agent must discover and refine. Quadruped robotics represents an embodied control domain where the agent must optimize both motor commands and the underlying objectives guiding behavior selection.
The system's applicability extends to domains where objective specifications are complex or poorly defined initially. In such contexts, human-designed solutions often suboptimize relative to evolved specifications. Hyperagents appears particularly valuable where domain experts lack clear criteria for evaluation, or where subtle interdependencies between objectives render manual specification infeasible.