The efficiency of agentic workflows is undergoing a fundamental shift. While traditional models load the entire conversation history at every reasoning step, Google has introduced SKILL.state, a new approach capable of slashing token consumption by 94% in long sessions without sacrificing accuracy.

Psst!Add AlexTech.ai asPreferred Source on Google

From Linear History to Structured State Representation

The core innovation lies in moving away from the "cascade memory" paradigm. Currently, AI agents maintain the full interaction history as input for subsequent reasoning, leading to an exponential increase in costs and latency. SKILL.state replaces this with a structured representation of the current state combined with the latest observation.

As the agent reasons through a problem, it actively writes only the information deemed useful for future steps into the state and discards the rest of the conversation history. This keeps the input size roughly constant regardless of session length. This evolution aligns with broader research into memory optimization, similar to the Always-On Memory Agent from Google Cloud, which aimed to simplify data management using SQLite.

Benchmarks: Gemini-3-Flash vs. Stateful Baselines

To validate the approach, Google conducted a 100-step benchmark using Gemini-3-Flash. The results show a stark difference in computational efficiency:

  • SKILL.state: achieved 0.94 accuracy using only 65k tokens.
  • Stateful baseline (LangGraph-style): achieved 0.91 accuracy, but required 1.1 million tokens.

The savings are massive: with nearly identical performance, the state-based system drastically reduces the token footprint. This directly addresses the "agentic multiplier" effect, where iterative workflows consume far more tokens than simple chat interactions—a trend previously noted when AI agents surpassed humans in token consumption on OpenRouter.

The Cognitive Prediction Constraint

Despite the impressive numbers, there is a significant caveat: the system relies on the agent's predictive capabilities. The effectiveness of SKILL.state depends entirely on the model's ability to identify which information will be needed in future steps. If the agent fails to write a crucial piece of data into the state, that information is permanently lost along with the discarded history, forcing a re-retrieval or leading to task failure.

This technical challenge highlights that token optimization is not just about compression, but about the intelligent management of relevance—a central theme for future autonomous agents operating on complex interfaces, as seen with Gemini 3.5 Flash's computer use capabilities.