AutoFyn Adapts Frozen Models for Long-Horizon Agents Through Persistent State

Summary

The AutoFyn technical report introduces an agent harness inspired by Expert Iteration that adapts a frozen model across repeated rounds without changing its weights. Each round starts with a fresh model session, while durable information is carried forward through explicit interfaces including memory files, reports, and repository state. An orchestrator uses specialized agents to explore, plan, and build multiple approaches, and a task-grounded verifier checks the result and produces an objective reward. That reward is distilled into persistent state, which changes the effective policy used in later rounds. The report formalizes the loop and its verification and state interfaces, then applies the system to olympiad mathematics, data science, and cybersecurity. On six fresh problems from the 2026 International Mathematical Olympiad, every tested model with room for improvement scored higher with AutoFyn than with its provider’s own coding agent. AutoFyn also produced the top-ranked agent on the Spider 2.0 dbt benchmark. In cybersecurity, it generated 16 maintainer-confirmed vulnerability advisories involving Next.js, MetaMask, pnpm, Warp, LiteLLM, Langflow, and Open WebUI. The results present persistent, reward-guided state as a way to improve long-horizon agent behavior while leaving the underlying model parameters unchanged.