Stateless Language Agents hit best results with 14.5x fewer tokens across 9 research tasks
RulinShao · x · 2026-10-13
Researchers from Stanford, CMU, UW and SambaNova (authors of ACE and Meta-Harness) introduce Stateless Language Agents (SLAs), built on the principle of "stateful search, stateless agents": no agent carries its conversation across invocations. Instead, an external harness owns the research state and reconstructs a fresh, role-specific context each call.
A stateless Advisor reads harness-summarized evidence and assigns experiments to parallel Workers, using under 0.6% of total tokens.
Across 9 tasks in software engineering, kernel optimization, and algorithm design with budgets up to 1B tokens:
- Best final result on every task
- Matches the strongest kernel baseline's final performance with 67.9M vs 986.3M tokens — a 14.5x (93.1%) reduction
- On FrontierSWE, SLA trails baselines at 25% of budget but leads on all four tasks at full budget
The paper traces long-horizon agent failures to growing histories, duplicated work, and token burn after progress stalls.
Related event: Stateless Language Agents Match Top Baselines with Far Fewer Tokens(2 posts)→
More from coding & agent
- Animating a cheetah via physics simulation reveals the limits of rigid-body skeletons — Kyrannio · 2026-10-13
- Running local AI agents filled a 1TB MacBook in three months — davidfromkansas · 2026-10-13
- 15 AI agents racing to notify me about the same email — omooretweets · 2026-10-13
- Every's guide to voice-driven agents: patch bugs and write by talking — every · 2026-10-13
- Delegating to your agent at 7am and going for a run: how do you count work hours now? — gethackteam · 2026-10-13
- Developer teaches Claude to paint in 3D, framing it as a pure-code motion-solving experiment — Kyrannio · 2026-10-13