Making Agents Stateless: A Better Route to Long-Horizon AI Research

Introduction

More inference does not automatically translate into more progress in automated research. When language agents operate for long periods, they may repeatedly replay an expanding conversation, duplicate experiments performed by other agents, or continue generating tokens after meaningful experimentation has stopped. These problems are partly architectural: where research state is stored and who decides what should happen next can matter as much as the underlying model.

In “Stateless Language Agents: Scaling Long-Horizon Automated Research,” a Stanford team proposes Stateless Language Agents, or SLAs. The central design principle is stateful search with stateless agents.

How the framework works

• Persistent state belongs to the harness. Candidate solutions, search directions, and measured outcomes are kept by the surrounding system instead of being carried through an agent’s conversation.
• Contexts are rebuilt for each invocation. Every call starts with a fresh, role-specific prompt. The harness decides which evidence an agent needs, rather than allowing an ever-growing transcript to determine visibility.
• Planning and execution are separated. A stateless Advisor reviews summarized evidence across search directions and assigns concrete experiments. Parallel Workers then execute those assignments and return measurable results.
• Evaluation extends beyond short horizons. The study compares SLA with three recent frameworks on software engineering, kernel optimization, and algorithm design, using budgets of up to one billion tokens.

Results and implications

According to the paper’s abstract, SLA achieves the best final result on every evaluated task. On kernel optimization, it reaches the final performance of the strongest baseline while using more than 84% fewer tokens. Ablations from shared checkpoints indicate that focused contexts and explicit assignments each contribute to progress, with their effects potentially compounding over longer runs. The Advisor consumes less than 0.6% of total tokens, suggesting that centralized coordination is not the dominant cost.

The broader lesson is that more conversational memory is not necessarily better research memory. For tasks driven by repeated experimentation, an explicit state store can preserve evidence without forcing every agent to reread the full history. Role-specific context can reduce distraction, while concrete assignments can limit duplicated work and keep parallel search directions productive.

The result also challenges short-horizon evaluation. Frameworks that appear similar when a benchmark saturates quickly may diverge once the budget becomes large enough for repetition, stagnation, and inefficient context use to emerge. In this view, a stateless agent is not an agent without memory. It is an agent whose durable memory has been moved from an implicit transcript into a structured, inspectable search process.

For future automated research systems, progress may depend not only on generating promising ideas, but also on recording evidence reliably, selecting the next experiment, and spending computation on directions that remain genuinely uncertain.

Hugging Face Daily Papers