TL;DR
This paper introduces Stateless Language Agents (SLAs), a framework that sustains progress in automated research over long horizons (up to one billion tokens) by separating research state from agent conversation. By using a central harness to manage state and providing agents with fresh, focused contexts for every task, SLAs achieve superior results while using up to 93% fewer tokens than existing stateful agent frameworks.
What Problem Does This Paper Address?
We are entering an era where Large Language Model (LLM) agents are tasked with "AutoResearch"—iteratively proposing, implementing, and evaluating solutions to complex engineering and scientific problems. However, the authors identify a critical "scaling wall": simply throwing more inference at a problem doesn't guarantee progress.
Current agent frameworks often rely on stateful agents that carry their entire conversation history across multiple steps. Over long horizons, this leads to two major failure modes:
- Context Rot: As histories grow, agents consume more tokens just to "remember" what they did, which eventually degrades their reasoning and leads to premature stopping.
- Coordination Collapse: Without explicit management, parallel agents often duplicate each other's work or get stuck in "experiment inactivity," where they stop making tool calls but continue to burn tokens.
Furthermore, the paper argues that current benchmarks (like circle packing) saturate too quickly, making it impossible to see how these systems perform when the search horizon extends to millions or billions of tokens.
Methods
The authors propose the SLA (Stateless Language Agents) framework, built on the principle of stateful search with stateless agents.
In this model, the "state" (the record of candidate solutions, scores, and evidence) lives in a system harness, not in the agents' memories. For every invocation, the harness reconstructs a fresh, role-specific context. The framework uses two distinct roles:
- The Advisor: A stateless agent that sees a summarized view of all search directions. Its job is global planning: deciding whether to diversify the search or concentrate on a promising bottleneck.
- The Workers: Stateless agents assigned to specific experiments. They work in isolated workspaces, seeing only their specific assignment and their own local progress.
Because every call starts with a "clean slate" context tailored by the harness, the agents avoid the baggage of long-running conversations while benefiting from a structured history of the entire research project.
Datasets
The authors evaluated SLA on tasks with significant "headroom"—problems where there is no known optimum and progress can continue for a long time:
- Software Engineering: Four tasks from FrontierSWE (e.g., converting libexpat to x86-64 Assembly, Git to Zig).
- Kernel Optimization: Anthropic’s VLIW SIMD task and NVIDIA’s SOL-ExecBench GPU tasks.
- Algorithm Design: The FrontierCS Structured-LWE cryptographic task.
Key Insights and Ideas
- Statelessness is a Feature, Not a Bug: By forcing agents to be stateless, the system prevents "context rot." The harness acts as a filter, ensuring agents only see what is relevant to their current task.
- Rankings Reverse Over Time: A crucial finding is that a method that looks best at 100 million tokens might be overtaken by another at 700 million tokens. Short-horizon evaluations can be deeply misleading for long-horizon research.
- Coordination is Cheap but Vital: The Advisor agent consumes less than 0.6% of the total tokens, yet its ability to prevent redundant work (Workers implementing the same feature) is essential for efficiency.
- Isolation Breeds Progress: Keeping Workers in isolated "lanes" allows different research directions to mature independently, preventing a single global "best" solution from prematurely killing off alternative approaches that might eventually prove superior.
Key Results
SLA achieved the best final result on every task evaluated, often with staggering efficiency gains:
- Token Efficiency: On the Anthropic kernel optimization task, SLA reached the strongest baseline's final performance using 14.5x fewer tokens (a 93.1% reduction).
- Performance Gains: On FrontierSWE, SLA improved scores by an average of 4.95 points over the strongest per-task baseline.
- Sustained Progress: Unlike baselines that plateaued, SLA continued to find improvements even as the budget approached one billion tokens.
Limitations and Future Work
- Cost of Evaluation: Long-horizon runs are expensive, limiting the number of independent trials the authors could run for some configurations.
- Executable Evaluators: The framework currently relies on tasks with automated, executable evaluators. Applying SLAs to more "vague" research goals (where evaluation is subjective) remains an open challenge.
- Advisor Quality: While the Advisor is cheap, its decision quality is the primary bottleneck. Future work could involve training specialized models for the Advisor role using the data generated by these long-horizon runs.
Summary updated