Stateless Language Agents: Scaling Long-Horizon Automated Research

Stateless Language Agents. Scaling Long-Horizon Automated Research. By Qizheng Zhang and seventeen others, from Stanford University, Carnegie Mellon University, and three other institutions. Published on the sixth of October. Its code repository is a placeholder: a readme, and nothing to run yet.

Automated research systems now run LLM agents over long horizons, and the hope is that more inference buys more progress. It does not, on its own. Agents replay histories that keep growing. They duplicate each other's work. Or they simply stop experimenting while tokens continue to be spent. These failures stay hidden because most evaluations use short budgets, or benchmarks that saturate early, so a run never lasts long enough for them to show.

The authors trace the failures to two design choices. Where does the research state live? And who decides what to try next? In a typical setup, the answer to the first is inside each agent's conversation, and that conversation only gets longer. The answer to the second is left to the agents themselves. Put those together and the three failures follow naturally. An agent rereading its whole past spends tokens on replay. Agents deciding independently can land on the same idea. And an agent with no one assigning it work can drift into spending without testing anything.

Stateless Language Agents, or SLAs, rest on one principle: stateful search with stateless agents. No agent carries its conversation from one invocation to the next. The harness owns the research state instead, meaning the candidate solutions and their measured outcomes, and it builds a fresh, role-specific context for every invocation. What an agent sees stops being whatever history the run happened to accumulate, and becomes an explicit design decision.

The SLA framework puts that principle into practice with two roles. A stateless Advisor reads evidence that the harness has summarized across the different search directions. It then assigns concrete experiments to Workers running in parallel. That division answers both of the earlier questions. Research state lives in the harness, as candidates and measured outcomes. Deciding what to try next belongs to the Advisor, which works from a summary rather than from a transcript. Nobody has to remember the run. The harness remembers it, and each call gets exactly the slice it needs.

SLA is evaluated against three recent frameworks on three kinds of task: software engineering, kernel optimization, and algorithm design, at budgets reaching one billion tokens. It achieves the best final result on every task. On kernel optimization, it reaches the final performance of the strongest baseline using over eighty-four percent fewer tokens.

Ablations run from shared checkpoints separate the two ingredients. Focused contexts and explicit assignments each contribute to progress, and their effects can compound over a full run. The Advisor itself is cheap: it uses less than zero point six percent of the tokens. The coordination, in other words, is not where the budget goes; nearly all of it is left for the Workers running experiments.

Two lessons come out, one about building these systems and one about judging them. Keeping durable research state out of agent conversations is what lets a long run keep making progress instead of drowning in its own history. And because the failures only appear over long horizons, short evaluations can misjudge both a research system and the parts it is made of.