Google researchers have introduced SKILL.state, an architecture capable of slashing token consumption in long-running agent workflows by up to 98%. Traditional agents built on the ReAct paradigm suffer from an inherent architectural flaw: at every single step, the model re-reads the entire history of its prior actions. Consequently, input prompt sizes scale linearly, cumulative inference costs explode quadratically, and the neural network inevitably begins to hallucinate as its reasoning degrades inside a bloated context window.
Google's engineering solution is pragmatic: instead of feeding an endless dialog transcript, the agent receives a compact, mutable execution state. The model generates an environment command alongside a concise diff patch containing updates relative to the previous step. The environment executes the command, and the LLM's next input contains only the updated state, the base system prompt, and the latest environment response. Context size remains fixed, turning exponential cost curves into strictly linear token growth.
In warehouse management benchmarks, this approach reduced token overhead by 93.8% across 100-step runs and 98% across 200 steps. In practical coding benchmarks like InterCode CTF, accuracy improved to 54.2% compared to the baseline 40–45%: the model stopped tripping over discarded hypotheses that previously cluttered its context. Crucially for enterprise production, the method requires zero model fine-tuning and works across any commercial API simply by modifying runtime prompting logic.