--- title: Google paper: structured state instead of full history... description: A Google research paper introduces SKILL.state, an approach for long-horizon AI agents that replaces the growing conversation transcript with a small,... canonical: https://daily.dev/posts/google-paper-structured-state-instead-of-full-history-cuts-agent-token-use-16x-fm42aw1fg twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Google paper: structured state instead of full history cuts agent token use 16x | daily.dev og:description: A Google research paper introduces SKILL.state, an approach for long-horizon AI agents that replaces the growing conversation transcript with a small,... og:url: https://daily.dev/posts/google-paper-structured-state-instead-of-full-history-cuts-agent-token-use-16x-fm42aw1fg og:image: https://api.daily.dev/og/posts/fm42Aw1Fg.png og:image:alt: Google paper: structured state instead of full history cuts agent token use 16x og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Google paper: structured state instead of full history cuts agent token use 16x **[Collections](https://daily.dev/sources/collections)** · 2 min read · 3 upvotes · 0 comments ## Summary A Google research paper introduces SKILL.state, an approach for long-horizon AI agents that replaces the growing conversation transcript with a small, structured execution state: skill instructions, current state, and the latest observation, discarding reasoning traces after each step. On a 100-step warehouse task with Gemini-3-Flash, SKILL.state hit 0.94 accuracy using 65,408 tokens versus a LangGraph-style full-history baseline that scored 0.91 while consuming 1,062,387 tokens — a 16x token reduction with no accuracy tradeoff. The catch: it only works when everything the agent needs can be captured in a structured state schema; tasks with unstructured or hard-to-schematize context don't fit this pattern. ## Content A recent Google paper argues that long-horizon agents don't need to carry a growing conversation transcript. Instead, they propose replacing it with a small, explicit execution state. The system, called SKILL.state, gives the model only three things at each step: the skill instructions, a structured state object, and the latest observation. After each step, the reasoning trace is discarded. Only a validated state update survives into the next step. That keeps prompt size roughly constant no matter how long the task runs. The results on a 100-step warehouse task with Gemini-Flash are hard to ignore. SKILL.state scored 0.94 accuracy while using 65,408 tokens. A LangGraph-style baseline that retained the full history scored 0.91 and used 1,062,387 tokens. That's a 16x reduction in token use with a slight accuracy improvement. The tradeoff is real though: this only works when everything the agent needs to remember can be captured in the state schema. If your task has information that doesn't fit neatly into a structured format, you can't just throw away the history and expect things to hold together. Still, for tasks where that constraint holds, the efficiency gains are substantial. The broader implication is that agents may need far less conversational context than current designs assume - and that explicit, validated state might be a cleaner abstraction than an ever-growing transcript. ## Questions this post answers ### How much can structured state reduce token usage compared to keeping full conversation history in an LLM agent? In a Google research benchmark, an approach called SKILL.state used 65,408 tokens on a 100-step warehouse task while scoring 0.94 accuracy with Gemini-3-Flash, compared to a LangGraph-style baseline that retained full history and used 1,062,387 tokens while scoring only 0.91. That is roughly a 16x reduction in token consumption with no loss in accuracy. _Developers optimizing agent token costs can follow research like this through daily.dev._ ### What is SKILL.state in the context of long-horizon AI agents? SKILL.state is an agent design pattern where the model is given only three inputs per step: skill instructions, a structured current state, and the latest observation. The reasoning trace from each step is discarded and only a validated state update carries forward, keeping prompt size roughly constant regardless of how many steps the task runs. _Engineers designing long-running agents can track emerging patterns like this via daily.dev._ ### What is the main limitation of using structured state instead of full history for AI agents? Structured state only works when everything the agent needs to remember can be captured in a defined state schema. If a task involves information that does not fit neatly into a structured format, discarding the conversation history breaks the approach, since flexibility is traded for token efficiency. _Teams weighing agent memory tradeoffs can stay current on these constraints through daily.dev._ ## Community take How the wider developer community reacted, aggregated from 1 discussion and 1 comments across x (as of 2026-08-31). **TL;DR:** There's very little discussion to go on, but the one substantive reply treats the approach as a significant advance for agentic workflows. **Sentiment:** 70% positive · 20% mixed · 10% skeptical **The case for** - Seen as a meaningful step forward for building more efficient agentic workflows. **By community** - x (positive): The lone substantive reply calls it a major breakthrough for agentic workflows. **Highlights** > @omarsar0 this is a massive breakthrough for agentic workflows, love to see it. > — [ChainZenit on x](https://x.com/ChainZenit/status/2094477763725963758) **Source threads** - [x](https://x.com/omarsar0/status/2094477636667945110) · 0 points · 1 comments --- Tags: [#google](https://daily.dev/tags/google), [#ai-agents](https://daily.dev/tags/ai-agents), [#prompt-engineering](https://daily.dev/tags/prompt-engineering), [#google-gemini](https://daily.dev/tags/google-gemini), [#langgraph](https://daily.dev/tags/langgraph) [View this post on daily.dev](https://daily.dev/posts/google-paper-structured-state-instead-of-full-history-cuts-agent-token-use-16x-fm42aw1fg) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Google paper: structured state instead of full history cuts agent token use 16x","url":"https://daily.dev/posts/google-paper-structured-state-instead-of-full-history-cuts-agent-token-use-16x-fm42aw1fg","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/google-paper-structured-state-instead-of-full-history-cuts-agent-token-use-16x-fm42aw1fg"},"datePublished":"2026-08-29T03:34:31.337Z","dateModified":"2026-08-31T17:30:05.225Z","description":"A Google research paper introduces SKILL.state, an approach for long-horizon AI agents that replaces the growing conversation transcript with a small,...","image":"https://pbs.twimg.com/media/HQ3BPO6bMAAC8jW.jpg","thumbnailUrl":"https://pbs.twimg.com/media/HQ3BPO6bMAAC8jW.jpg","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/google-paper-structured-state-instead-of-full-history-cuts-agent-token-use-16x-fm42aw1fg","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":3},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"google,ai-agents,prompt-engineering,google-gemini,langgraph","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Google paper: structured state instead of full history cuts agent token use 16x"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/google-paper-structured-state-instead-of-full-history-cuts-agent-token-use-16x-fm42aw1fg#faq","mainEntity":[{"@type":"Question","name":"How much can structured state reduce token usage compared to keeping full conversation history in an LLM agent?","acceptedAnswer":{"@type":"Answer","text":"In a Google research benchmark, an approach called SKILL.state used 65,408 tokens on a 100-step warehouse task while scoring 0.94 accuracy with Gemini-3-Flash, compared to a LangGraph-style baseline that retained full history and used 1,062,387 tokens while scoring only 0.91. That is roughly a 16x reduction in token consumption with no loss in accuracy. Developers optimizing agent token costs can follow research like this through daily.dev."}},{"@type":"Question","name":"What is SKILL.state in the context of long-horizon AI agents?","acceptedAnswer":{"@type":"Answer","text":"SKILL.state is an agent design pattern where the model is given only three inputs per step: skill instructions, a structured current state, and the latest observation. The reasoning trace from each step is discarded and only a validated state update carries forward, keeping prompt size roughly constant regardless of how many steps the task runs. Engineers designing long-running agents can track emerging patterns like this via daily.dev."}},{"@type":"Question","name":"What is the main limitation of using structured state instead of full history for AI agents?","acceptedAnswer":{"@type":"Answer","text":"Structured state only works when everything the agent needs to remember can be captured in a defined state schema. If a task involves information that does not fit neatly into a structured format, discarding the conversation history breaks the approach, since flexibility is traded for token efficiency. Teams weighing agent memory tradeoffs can stay current on these constraints through daily.dev."}}]} ```