代理架構與評測
SKILL.state Discards Agent Execution History, Cutting Token Use for 100-Step Tasks to One-Sixteenth With Structured State
Researchers from Google and Purdue University designed an agent that reads only a fixed skill specification, its current state, and the latest observation at each step, immediately discarding its reasoning history after validating state updates. In a 100-step warehouse test, it achieved 0.94 accuracy using 65,408 tokens, but its effectiveness depends on a predefined state structure preserving information needed later.
SKILL.state attempts to transform long-running agents from “ever-growing conversations” into runtimes that more closely resemble traditional state machines. At each step, the input contains only an immutable skill specification, the structured current state, and the latest observation from the environment. After the model proposes a state patch and the next action, a deterministic runtime validates the patch against a schema and commits it only if validation succeeds; all previous natural-language reasoning and tool history are then immediately discarded. Formatting errors trigger a rollback and retry, preventing the model from directly corrupting persistent state.
In synthetic warehouse tasks using Gemini-3-Flash, SKILL.state maintained an average prompt length of roughly 1,736 to 1,905 tokens as tasks scaled from 10 to 200 steps. In the 100-step test, it achieved 0.94 accuracy while consuming 65,408 tokens in total; a LangGraph-style baseline that retained history achieved 0.91 accuracy while using 1,062,387 tokens—a 16.2-fold difference. At 200 steps, SKILL.state still achieved 0.94 accuracy using approximately 122,000 tokens, while the summary-memory baseline fell to 0.84 and consumed about 6.18 million tokens. The researchers also ran controlled experiments with a comparable 1,800-token budget: a sliding window, capped summaries, and LLMLingua achieved only 0.18, 0.52, and 0.22, respectively. This suggests that the advantage comes not merely from shortening prompts, but from explicitly preserving operational state such as item locations and dependencies.
Public interactive benchmarks showed results in the same direction: pass@1 on InterCode CTF reached 54.2%, compared with 46.4% for the best baseline; on the τ-Bench retail and airline tasks, it scored 58.3% and 32.4%, respectively. However, the state schema is predefined by humans. If the model fails to anticipate that a piece of information will be useful later, it may be impossible to recover once the history has been deleted. All methods may also fail when the external world changes in ways not captured by the observed event descriptions. The paper makes its prompts public but does not provide a complete, independently maintained library. Future work should test more open-ended software engineering and browser tasks, semantic validation of state transitions, and the latency introduced by schema design and retries.