September 2026 Summaries

AI agents are emerging as a distinct audience for company documentation, using structured machine-readable resources such as Markdown pages, llms.txt files, and Model Context Protocol tools rather than browsing like humans. Mintlify’s traffic and survey data indicate that agent and documentation-assistant usage is growing substantially faster than human readership, while controlled tests found that pairing Markdown with an llms.txt navigation map sharply reduced agent 404 errors and token consumption compared with HTML alone. The report argues that documentation quality, structure, and currency increasingly affect product adoption because agents may act on inaccurate information without the skepticism human users apply. Although many teams now use AI agents to draft documentation updates, most retain human review, and organizations continue to struggle with fragmented knowledge spread across documentation, support, community, sales, and product systems. Survey findings suggest few companies have made all knowledge surfaces accessible to agents or assigned clear ownership for doing so, prompting a growing emphasis on “knowledge engineers” and unified knowledge infrastructure that can keep content synchronized, discoverable, and continuously improved for both humans and AI agents.

AI agents are becoming a major audience for company knowledge, accounting for 66% of measured traffic on Mintlify-powered documentation in July 2026, which increases the importance of accurate, structured, and retrievable information. The article argues that organizations need “knowledge engineers” to govern the systems behind internal and external knowledge, reducing risks such as agents retrieving outdated documentation, conflicting policies, or incorrect product information that can spread through automated support, generated code, and workflows. This role combines content strategy, technical infrastructure, and organizational governance by establishing authoritative sources, linking product and support changes to documentation updates, publishing information in agent-accessible formats such as Markdown, llms.txt, APIs, and MCP tools, and monitoring unanswered questions or retrieval failures. While agents require detailed specifications, schemas, and procedures, human readers benefit more from mental models and guidance about choices and trade-offs. A knowledge engineering approach can accelerate updates, expose content gaps through agent behavior, and allow experts to focus on review and higher-value work, but it requires clear accountability for maintaining trustworthy knowledge across the organization.

Mintlify’s 2026 State of Knowledge Report finds that AI agents have become a major audience for company knowledge, generating 257 million requests across Mintlify-powered sites in August 2026 versus 131 million human page loads. The report argues that agents should not be treated like indiscriminate bots because they are typically acting on users’ behalf, with most traffic arriving through machine-readable resources such as Markdown pages, llms.txt files, and agent skills. Controlled navigation tests across 20 documentation sites showed that linking Markdown pages to an llms.txt content map substantially reduced failed requests and token consumption, while documentation-update workflows are increasingly agent-assisted: nearly 367,000 update requests were made through Mintlify between February and August, though humans usually retain publishing oversight. The report also highlights that company knowledge extends beyond formal documentation to support, forums, changelogs, and internal or customer resources, often scattered across teams and tools, leaving many organizations less prepared for agent access than they believe.

Mintlify has introduced fixed, outcome-based pricing for its AI features, replacing token-based charges that varied by run and were difficult for customers to predict. Assistant responses now cost 25 credits only when they successfully answer a question, while documentation automations cost 250 credits only when they make an update; unsuccessful answers and checks that find no needed changes are free. The company estimates that 96% of teams will use the same or fewer credits for the assistant, more than 99% will do so for automations, and automation users may spend about 70% fewer credits because most no-change runs will no longer be charged. Pro and Enterprise plans now also include unlimited editor and Slack agent usage, while existing plans, included credits, prepaid credits, and per-credit prices remain unchanged. The new model applies automatically and is intended to make AI costs predictable by charging for delivered results rather than computational effort.

Mintlify has redesigned its llms.txt generation system to prevent oversized files from overwhelming AI agents with irrelevant documentation, which can increase token costs and slow retrieval. Instead of using a single file capped at 100,000 characters, the new architecture creates a navigation-based hierarchy in which sections such as guides, API references, and integrations are moved into descriptive child files under /_llms/, allowing agents to follow only relevant paths. The generator uses a site’s docs.json navigation structure to identify logical groupings and repeatedly moves the sections that save the most space until each file fits the character limit. In tests across 200 retrieval tasks on ten documentation sites using Claude Sonnet 5 and GPT-5.6 Terra, the hierarchical approach reduced average task completion time by 52.1%, token use by 44.8%, and agent 404 errors by 88.1%. The feature is now available to all Mintlify customers and regenerates automatically whenever documentation is published.