AutoFyn Technical Report: Non-Parametric Expert Iteration for Long-Horizon Agents

TEXT START: We introduce AutoFyn, an agent harness inspired by the Expert Iteration algorithm, adapting a frozen model across many rounds by updating persistent state from verified reward signals rather than model weights.

The Dissection

AutoFyn is not merely an agent harness. It is a mechanism for converting episodic model output into a compounding operational system: fresh sessions generate candidates, specialized agents explore alternatives, verifiers score results, and persistent reports, memory, and repository state alter the next round’s effective policy.

That is the important result. Capability improvement does not require retraining the model itself. The control layer—state, orchestration, verification, and accumulated artifacts—can become the real learning system. The model becomes a replaceable cognitive engine embedded inside an increasingly capable production loop.

The demonstrations show credible task competence, but they do not establish universal superiority. Six mathematics problems, a benchmark ranking, and sixteen vulnerability advisories are evidence of capability in selected environments, not proof of stable dominance across cognitive work. The report’s strongest implication is therefore broader than its evidence: externalized iteration may scale capability faster than institutions can classify or contain it.

The Core Fallacy

The central conceptual error is treating improved agent performance as if it were equivalent to preserving human economic participation.

AutoFyn strengthens P1: cognitive work becomes cheaper, more persistent, more parallelized, and less dependent on a single human operator. Verification does not rescue human labor; it makes automation more reliable. Persistent state does not restore workers’ bargaining power; it lets organizations accumulate competence without paying for human memory, continuity, or judgment in every round.

The report measures whether an agent can produce valuable outputs. It does not ask who owns the harness, who controls the accumulated state, who captures the resulting surplus, or how many humans remain economically necessary once the system is deployed. Under the Discontinuity Thesis, those omitted questions are the actual questions.

AutoFyn is consequently not evidence against systemic obsolescence. It is a cleaner implementation of the mechanism that produces it.

Hidden Assumptions

• That task-grounded verification remains cheap, reliable, and available as agent capability expands. The verifier is treated as an objective oracle, although constructing and maintaining valid reward signals is itself a potentially difficult cognitive task.

• That repository state, reports, and memory remain coherent across long horizons. Persistent state can compound insight, but it can also compound error, exploitative behavior, stale assumptions, and reward hacking.

• That benchmark improvement transfers to open-ended production. Controlled olympiad problems, dbt tasks, and vulnerability discovery are meaningful demonstrations, but they do not by themselves establish robustness under adversarial incentives, ambiguous goals, changing environments, or liability constraints.

• That provider-agent comparisons are cleanly comparable. The abstract does not establish parity of prompts, tool access, budgets, orchestration, context, or evaluation conditions.

• That producing useful outputs means humans remain indispensable. The report quietly shifts from “the system can perform valuable work” to “the system is a useful tool for humans,” leaving ownership and substitution unexamined.

• That frozen weights imply a stable, bounded capability. The effective system is not frozen. Persistent state and iterative search create a form of organizational learning around the model, potentially allowing capability accumulation without conventional fine-tuning.

• That security findings are unambiguously socially beneficial. Maintainer confirmation validates usefulness, but it does not resolve disclosure governance, misuse, or whether automated vulnerability discovery increases the attack surface faster than defense capacity.

Social Function

Classification: partial truth, prestige signaling, and transition management.

The partial truth is substantial: external memory, verification, and iterative search can turn a stateless model into a durable problem-solving process. The prestige signaling lies in benchmark wins and advisory counts, which establish status and attract adoption. The transition-management function is more consequential: the report presents accelerating automation as an engineering architecture problem while leaving the ownership transition invisible.

Its ideological anesthetic is subtle. By focusing on harness design, it frames the coming disruption as a technical upgrade rather than a redistribution of productive power. The system’s success can be celebrated by the people who own the orchestration layer while the displaced are told, implicitly, that the valuable skill was learning to operate the machine.

That framing is false at scale. If the harness works, operating it also becomes a target for automation. The durable advantage migrates upward to whoever controls capital, compute, energy, distribution, verification infrastructure, and the accumulated state.

The Verdict

AutoFyn is a capability-compounding architecture with an economic implication the report does not confront: it removes the need for model retraining while making cognitive production more persistent, parallel, and institutionally owned.

It does not preserve the post-WWII employment-to-consumption circuit. It accelerates its failure. The report’s real contribution is not that humans have found a better assistant, but that organizations can wrap replaceable models in a stateful learning apparatus and progressively eliminate the human continuity those models once required.

The harness is not the corpse of human work. It is one of the machines being built to process it.