$ cat wiki/papers/2026/2610.07625-stateless-agents.md
Stateless Language Agents — durable research state outside conversations
TL;DR
SLA keeps research state in the harness and starts each agent invocation with a fresh, role-specific context. The authors report stronger long-run research outcomes than three recent frameworks in their evaluations. (source)
Authors & Org
Qizheng Zhang, Changxiu Ji, Isaac Sun, Yuetai Li, Shubhangi Upasani, Sherry Ruan, Boyuan Ma, Fenglu Hong, Vamsidhar Kamanuru, Yoonho Lee, Yuzhen Mao, Genghan Zhang, Rulin Shao, Qiuyang Mang, Andy Dimnaku, Changran Hu, Radha Poovendran, Kunle Olukotun. Submitted 2026-10-06. Affiliations are not supplied by the abstract page and are not inferred from author names. (source)
Method
The harness retains candidate solutions and measured outcomes. A stateless Advisor reads summarized evidence across search directions and assigns concrete experiments to parallel Workers; no agent carries its conversation between invocations. Evaluations cover software engineering, kernel optimization and algorithm design at budgets up to one billion tokens. (source)
Results
The authors report the best final result on every evaluated task, reaching the strongest kernel baseline's final performance with over 84% fewer tokens. The Advisor consumes less than 0.6% of tokens. Shared-checkpoint ablations find contributions from both focused contexts and explicit assignments, with effects that can compound over full runs. (source)
Significance
For Agents (LLM Agents), the design makes persistence a harness responsibility. For R&D Automation Index, it argues that short evaluations can misjudge systems that continue experimenting over long budgets. These are interpretations of the paper's design and reported experiments. (source)
Open Questions
The abstract does not give the complete task list, named baseline configurations or monetary costs. Full-paper and code review are needed to establish reproducibility and whether the reported token savings generalize beyond the tested kernel workloads. (source)
Cite
arXiv:2610.07625, submitted 2026-10-06. Captured abstract.