Stateless Language Agents: Scaling Long-Horizon Automated Research
The paper argues long-horizon research agents scale better when memory lives in the harness, not in agent chat histories.
Its Stateless Language Agents reset each agent between invocations while a central system tracks candidates, results, and evidence. An Advisor uses that state to assign experiments to parallel Workers. The authors report best final results across software engineering, kernel optimization, and algorithm design tasks, with one kernel baseline matched using over 84% fewer tokens. They also say short-budget evaluations can miss the failure modes that show up at billion-token runs.
HF Daily Papers' note
Its Stateless Language Agents reset each agent between invocations while a central system tracks candidates, results, and evidence. An Advisor uses that state to assign experiments to parallel Workers. The authors report best final results across software engineering, kernel optimization, and algorithm design tasks, with one kernel baseline matched using over 84% fewer tokens. They also say short-budget evaluations can miss the failure modes that show up at billion-token runs.
HF Daily Papers' note
score 5