Stateless language agents cut token use over 84% in long-horizon research tasks

Long-horizon LLM agents often consume tokens so quickly that billion-step research becomes impractical, because each call replays an ever-growing history of intermediate results. Stateless Language Agents (SLAs) address this by moving durable state into an external harness, so each worker sees only the concise evidence it needs. According to the paper, SLAs match the strongest kernel baseline's final performance while using over 84% fewer tokens, across software engineering, kernel optimization, and algorithm design tasks at budgets up to a billion tokens. The coordinating Advisor component accounts for under 0.6% of token usage. The approach depends on the harness summarizing evidence accurately, and the evaluation covers only three engineered domains, leaving its applicability to less structured research open.