AutoFyn is an agent framework inspired by the Expert Iteration algorithm. Each round starts with a fresh model session, yet persistent information such as memory files, reports, and repository state is re‑introduced through explicit interfaces, allowing verified reward signals to update the persistent state and indirectly shape the policy for the next round.
Within a round, an orchestrator explores, plans, and generates many alternative approaches while specialized sub‑agents carry out concrete tasks. A task‑grounded verifier evaluates these approaches, provides an objective reward, and distills that reward back into the persistent state. The loop has been validated in three domains: olympiad mathematics, data science, and cybersecurity.
On the six new problems of the 2026 International Mathematical Olympiad, every model with room for improvement scored higher under AutoFyn than its original coding agent. AutoFyn also built the top‑ranked agent on the Spider 2.0 dbt benchmark and produced sixteen maintainer‑confirmed vulnerability advisories across Next.js, MetaMask, pnpm, Warp, LiteLLM, Langflow, and Open WebUI.
Review: AutoFyn shows promise in leveraging persistent state and reward distillation to improve across iterations, achieving notable results in mathematics, data science, and cybersecurity.