Agon: An Autonomous Large-Scale Omnidisciplinary Research System Built on Prompt Economy

Agon is an autonomous research orchestrator designed to shift the bottleneck of scientific production from artifact creation to the evaluation of claims. By implementing a series of "factories"—specialized loops for ideas, proposals, experiments, and papers—the system automates the research lifecycle. The core philosophy is "machine scales, human steers," where the system handles the heavy lifting of execution and iteration while human researchers provide the high-level direction and final judgment.

The system is built on six foundational principles designed to maximize the return on investment for prompt engineering and ensure long-term maintainability:

• Prompt Economy & Minimal Prompts: Focuses on creating reusable, short, and efficient agent loops rather than task-specific prompts that are costly to maintain.
• Future-Facing: Prompts are designed to be model-agnostic, ensuring the system improves as underlying LLMs advance without requiring constant rewrites.
• OmniDisciplinary & Massive Parallelism: The architecture is domain-agnostic, allowing it to run thousands of concurrent research threads across diverse scientific fields.
• Zero-Code Orchestration: Research workflows are managed via natural-language prompt dispatchers rather than brittle, hard-coded state machines.

Agon operates through an adversarial producer-critic model. Each research artifact is generated by an agent and then subjected to an independent audit by a critic agent, which attempts to break or invalidate the work. This ensures that only high-quality outputs advance through the pipeline.

Because autonomous systems are prone to specific failure modes, the authors developed a taxonomy to categorize these issues by severity, fixability, and visibility. This taxonomy helps distinguish between "perception failures" (where the model misinterprets data), "reasoning failures" (logical errors), "execution failures" (code or tool errors), and "motivation failures" (lack of scientific significance). By mapping these failures, the system identifies which tasks can be automatically corrected by the loop and which require human intervention.