“Determine whether your software is needed; the Agent only requires 500 Tokens.”
Google proposes an Agent engine optimization framework, requiring software documentation to clarify core information within the first 500 Tokens and limit the total length.
2026-09-07 18:32Products & Apps🔥 47.2 heat score
2sources
1days unfolding
47.2heat score
6mentions
SummaryAI generated
On September 7, 2026, Addy Osmani, Director of AI Engineering at Google Cloud, introduced the concept of Agent Engine Optimization (AEO) aimed at addressing the issue where the current Coding Agent ignored lengthy documents or fabricated non-existent interfaces due to context window limitations. Osmani noted that some software had excessive document length, redundant formatting, or too many HTML wrapping elements, preventing the Agent from effectively invoking them. To address this, he recommended that developers implement six optimization measures: check robots.txt to ensure AI crawlers are not blocked; publish an llms.txt file with the number of tokens to accurately identify document structure; write skill.md containing core capabilities, input parameters, and constraints as the product’s “self-recommendation letter”; provide an accessible Markdown version to reduce Token consumption; implement a block-based strategy for handling long documents. Additionally, Osmani emphasized that developers needed to manage the first 500…
Google proposes an Agent engine optimization framework, requiring software documentation to clarify core information within the first 500 Tokens and limit the total length.
Coverage · reports per dayLANGUAGE SPLIT
Entity relations
Integrated timelineUNIFIED TIMELINE
2026-09-07
Google proposes core requirements for the AEO framework
Addy Osmani, Engineering Director of Google Cloud AI, proposes the concept of Agent engine optimization, indicating that some software fails to allow effective invocation of agents due to overly long documents or redundant formats. Developers are advised to clarify what it is, what it can do, and how to… within the first 500 Tokens of the documentation.
Addy Osmani, Engineering Director of Google Cloud AI, introduced the concept of “Agent Engine Optimization (AEO)”, stating that developers should clearly explain “what it is, what it can do, and how to get started” within the first 500 Tokens of the document, and control the quick-start guide to be less than 15,000 Tokens and the API reference page to be less than 25,000 Tokens. Osmani recommended resolving misblocking of AI crawlers by auditing robots.txt, publishing an llms.txt file with the number of labeled Tokens to guide the Agent in accurately locating documents, and creating a skill.md document containing core capabilities, input parameters, and restrictions as the product’s “self-recommendation letter”. It was also necessary to ensure that the document was available in a pure Markdown format for direct access to reduce Token consumption. These measures aim to address the issue where the current Coding Agent ignores long documents or fabricates interfaces due to context window limitations, preventing the product from being…
Addy Osmani, Engineering Director of Google Cloud AI, introduced the concept of AEO (Agent Engine Optimization). He pointed out that some software fails to allow effective invocation of agents due to overly long documents or redundant formats, and even creates non-existent interfaces. The key difference is that the agent only reads the first 500 Tokens and strictly limits the context window, ignoring HTML wrapping elements and lengthy navigation. To address this, Osmani proposed a six-layer optimization framework: checking robots.txt to ensure that AI crawlers are not blocked; publishing llms.txt files to accurately determine the document structure; writing skill.md to declare the core capabilities and constraints of the product; providing an accessible Markdown version to reduce Token consumption; and implementing a chunking strategy for handling long documents. These practices aim to enable the agent to preferentially select and efficiently use the target software.