Google Paper Slashes Agent Token Use by 94% with State Tracking

A new paper from Google introduces SKILL.state, a method designed to drastically reduce the token usage of AI agents during long interactions. Traditional agents maintain extensive conversation histories as part of their input, leading to escalating token counts and costs. SKILL.state proposes a shift from this history-based approach to a structured representation of the agent's current state and the latest observation.

During its reasoning process, the agent identifies and stores crucial information for future steps within this state representation, effectively discarding the lengthy conversation history. This allows the input size to remain relatively constant, even over extended sessions. Benchmarks using Gemini-3-Flash demonstrated that SKILL.state achieved 0.94 accuracy with only 65,000 tokens, a significant improvement over a stateful baseline that required 1.1 million tokens for comparable accuracy.