Google Introduces SKILL.state Architecture to Save Tokens in Agent Sessions

Google Introduces SKILL.state Architecture to Save Tokens in Agent Sessions

The authors propose the SKILL.state architecture: instead of accumulating dialogue history, it uses an explicit, mutable structured execution state.

Typically, agents operate on a ReAct-style principle, meaning that at each step, the entire accumulated dialogue is fed back into the model. This causes the prompt to grow linearly each time, while the total token consumption for the entire run grows quadratically. Additionally, the model is forced to extract the current task state from a mountain of text each time, which can sometimes provoke hallucinations.

The article suggests instead sending the model an explicit mutable structured execution state (referred to as execution state in the article) at each step. Here’s how it works:

– At each step, the model, in addition to reasoning, outputs a JSON block with two fields: (1) a specific command that the agent must execute in the environment; (2) state_patch – information about what has changed relative to the previous state.

– The agent executes the command and receives a response from the environment, and the current state is updated according to the state_patch.

– For the next step, the updated state, system prompt, and the last response from the environment are passed. No previous messages, model reasoning from past steps, or anything else.

Thus, the prompt size at each step becomes constant, and the total token costs grow linearly.

In a synthetic warehouse management benchmark, SKILL.state saves 93.8% of tokens at 100 steps, and 98% at 200 steps. The longer the context, the more dramatic the gap.

In real benchmarks, the savings are slightly lower, but SKILL.state sometimes even wins in accuracy. For example, on the InterCode CTF, it achieves 54.2% compared to 40-45% for baselines. This is due to the fact that representing the task through state results in a cleaner and less cluttered view free from distracting facts and incorrect hypotheses from past contexts.

This solution can be applied to any model, even closed ones, simply by changing the runtime. Appendix A provides the necessary prompt template, so it can be taken and used.