For AI agents: a documentation index is available at /llms.txt — markdown versions of all pages are available by appending index.md to any URL path.
AI & Agent Research

AI & Agent Research

I research how AI agents work, how they consume information, and how the ecosystems around them are evolving. This page collects that work in one place. For my documentation background, see Documentation & Developer Education. For my programming projects, see Programming.

Talks & Interviews

  • State of Docs Report 2026 - Featured discussing AI consumption of documentation. Podcast forthcoming.
  • Can ANY AI Pass This Agent Reading Test? (BetterStack YouTube, 2026) - Walkthrough of the Agent Reading Test, a benchmark for measuring how AI agent web fetch pipelines handle documentation failure modes.
  • Why AI Agents Struggle with Modern Documentation (YouTube, 2026) - Interview covering how agents access documentation in real time and the failure modes most docs teams don't know about.
  • Designing Documentation for Agents, Not Just Users (Behind the Docs podcast, March 2026) - How AI agents struggle to use existing documentation, common structural issues like truncation and hidden content, and the Agent-Friendly Documentation Spec as a solution framework.
  • When AI Reads the Docs: LLMs, Agents, and Documentation Design (Deborah Emeni's YouTube Coffee Chat, March 2026) - How LLMs and agents are two very different documentation consumers, what "AI-friendly documentation" actually means, how documentation structure affects machine interpretation, and how documentation practices may evolve as machines increasingly consume technical content.

Specifications & Standards

Agent-Friendly Documentation Spec

A specification defining 23 checks across 7 categories for evaluating how well a documentation site serves AI agent consumers. Covers llms.txt discovery, markdown availability, page size, content structure, URL stability, and more. Based on real-world agent access patterns I've been researching since late 2025.

Tools

afdocs

A CLI tool that implements the Agent-Friendly Documentation Spec and tests docs sites against it. Point it at a URL and it reports where your docs stand. Published on npm. Fern's Agent Score directory uses afdocs to score API documentation sites at scale.

agentsummons

A Go library and CLI for invoking agent harnesses (Antigravity CLI, Claude Code, Codex CLI) in headless mode. Every headless-agent experiment rediscovers the same lore: which binary, which permission-bypass flag, which harness swallows your prompt. agentsummons owns that knowledge once, behind one API and one CLI.

agentminutes

Meeting minutes for your agents: a Go library and CLI that parses native agent harness session transcripts into one unified, comparable event schema. Transcripts are ground truth for how an agent actually behaved (which tools it chose, what it saw back, what it spent); agentminutes makes them analyzable and comparable across harnesses. Companion to agentsummons: agentsummons convenes the meeting, agentminutes takes the minutes.

skill-validator

A CLI that validates Agent Skills against the agentskills.io specification. Checks directory structure, frontmatter, content quality, cross-contamination risk, and token budget composition.

skillxp

A skill invocation runner: installs an Agent Skill in a fresh fixture, invokes an agent harness headlessly, and reports what actually reached the model, with transcript evidence rather than model self-reporting. Builds on agentsummons and agentminutes, and powers the agent-skill-implementation loading benchmark.

Research & Analysis

Agent Skill Ecosystem Analysis

An ecosystem-scale analysis of 673 Agent Skills across 41 repositories, examining compliance with the Agent Skills specification and content quality. Includes an interactive dashboard and a downloadable paper.

Agent Skill Implementation Research

Empirical research into how agent platforms actually implement Agent Skill loading, management, and presentation. Catalogs 40 checks across 10 categories, with 33 benchmark skills containing canary phrases for testing platform behavior without relying on model self-reporting. An automated benchmark runner generates reports for Antigravity CLI, Claude Code, and Codex CLI.

Agent Reading Test

A benchmark for measuring how AI agent web fetch pipelines handle real-world documentation failure modes. 10 test pages target specific failures (truncation, SPA shells, tabbed content, redirects, soft 404s) using canary tokens embedded at strategic positions. Task-first design prevents relevance-layer priming, and human-side scoring avoids agent self-report inflation.

Agent Web Fetch Behavior

Research into how coding agents actually fetch and process web content, including truncation behavior, redirect handling, and content negotiation across platforms.

Agent-Friendly Documentation Audit

An analysis of hundreds of documentation pages across popular developer tools, examining how well they serve AI agent consumers. The research that led to the Agent-Friendly Documentation Spec.

Writing

I write about agents, documentation, and the AI ecosystem on this blog and at AE Shift.

Selected articles:

For all AI-related posts, see the ai tag.