Where Autonomy Belongs: A Separation of Powers for Agentic Program Evolution
Summary
- LLM-driven program evolution relies on three decisions: govern, propose, and judge, which existing architectures often misallocate.
- The authors introduce EvoTrias, a separation-of-powers architecture assigning each decision to a distinct owner communicating via typed contracts.
- EvoTrias features Structured Governance, Open Proposal, and Protected Judgment to improve reliability and efficiency.
- Separating Governance from Proposal reduces token costs by 64%, while Open Proposal scales evolution to a 680K-line industrial repository.
- Protected Judgment preserves benchmark reliability, and EvoTrias establishes new state-of-the-art results on mathematical, systems, and algorithm benchmarks.
Why it matters
Proper allocation of autonomy in agentic program evolution architectures can significantly reduce computational overhead, lower token costs, and maintain benchmark reliability when scaling to large codebases.