Agents on Dachary Carey

What Changed in Agent Traffic?

Sun, 18 Oct 2026 12:00:00 -0500

When I updated my agent traffic classifier last month, I wrote about some observations related to agent data seeking patterns. And one of my takeaways was that five months was too long to let a traffic classifier sit in the fast-paced world of AI development, so today I did another check in to see how things are going. In the month since my last check, I've observed more changes, so let's talk about what's moving in agent traffic and what you need to care about if you're watching this.

Agent Data Seeking Patterns

Wed, 23 Sep 2026 07:00:00 -0500

Two conversations this month sent me back into my server logs. The first was the ongoing industry debate about whether llms.txt does anything, which flares up every time a well-known project adds one or removes one. The second was a colleague mentioning that answer engines and crawlers were fetching an llms-info file from their site, which was news to me, because I'd never seen a request for one.

Both are really the same question: what do agents actually reach for when they visit a website? Not what we've told them to reach for, not what the spec du jour says they should reach for, but what shows up in the logs when nobody is watching. I've been running an agent traffic classifier and signal tracker across my sites since April, and it had been a while since I validated the patterns baked into it. So I pulled nine days of logs (September 11 through 19) from twelve sites, re-classified everything, and went looking.

What "Just Prompting" Looks Like

Sun, 20 Sep 2026 11:00:00 -0500

A colleague asked me this week how I've built the things I've built over the past year: the Agent-Friendly Documentation Spec, the afdocs checker, agentsummons and agentminutes, Agent Skill Implementation, the Agent Reading Test. He characterized it as "just prompting," and then asked whether there was something in how I prompt that we could teach non-technical colleagues in his department.

I was skeptical. My private theory was that whatever works for me works because I have enough domain background to notice when the agent is wrong, and that isn't something you can easily hand off. But a private theory is exactly the kind of thing I'd tell an agent to go verify, so I did the same thing to myself. I had Claude Code read every prompt I've given it since December 2025 and tell me what it found.

Does Your Agent Know What It Can Do?

Tue, 21 Jul 2026 07:00:00 -0500

I've been building some new tools lately, and have noticed some interesting patterns in Claude Code around discovering its own capabilities and limitations. It reminded me of some conversations I've been having with fellow AI researcher, Rhyannon Rodriguez, about how agent harnesses often lack introspection/awareness and debugging around their own capabilities. I think there are interesting observations here for agent harness developers, and open questions for the rest of us, about how agents understand our products.

A Skill is More than Markdown

Mon, 11 May 2026 07:00:00 -0500

Organizations have been pushing official company skills out the door because skills look like just a markdown file, and that seems simple and trivial. But when you dig into the lifecycle of an official skill that you intend to support, distribute, and stand behind, it's a lot more than a markdown file. I'm not sure a single organization shipping skills today recognizes how much more.

This is the article I promised at the end of Can Agent Skills Make Output Worse?, where I closed by saying I'd been ideating around what skill lifecycle management looks like and would write more about it as I proposed process and tooling around it. This is the lifecycle that I'm picturing. Let me tell you what's actually involved, where the work falls, and where the org chart runs out of people to do it.

Can Agent Skills Make Output Worse?

Thu, 07 May 2026 22:00:00 -0500

In the last ~6 months, Agent Skills have gone from a new feature introduced by Claude Code to a widely adopted paradigm across the industry. Companies that want to signal they understand customer AI adoption and consider it as part of the developer experience have been publishing official skills. I have been part of such an effort at my own company, where I have learned a ton from our experience and from analyzing other releases across the industry. I went in expecting a documentation-pipeline problem. I came out understanding that official skills are a full product surface; one that can actively make your customers' agent output worse if you ship them naively.

Astro Removed its llms.txt

Mon, 04 May 2026 07:00:00 -0500

As I have been re-scoring documentation sites with my updated afdocs tool, I have had the opportunity to see a snapshot of various docs sites a few weeks apart as the industry is starting to pay a lot more attention to how to make documentation agent-friendly. I got a surprise when I re-scored the docs for Astro, a web framework that is well-regarded. In the prior scoring run, Astro got a C (78) for agent-friendliness; their new score is an F (59); capped by a missing llms.txt. I wanted to dig into why someone would make a change to make their docs less agent-friendly, when you have a contrast like last week's GitHub going out of its way to build their llms.txt into a comprehensive agent-discovery layer.

GitHub Reimagined llms.txt as an API

Fri, 01 May 2026 07:00:00 -0500

I'm working on a new research report that starts by scoring a wide cross-section of documentation sites with afdocs. Consider it an industry-wide audit driven by specific research questions. This has given me an excuse to dig into different platforms, web hosts, documentation sizes, and documentation strategies. I've really been surprised and intrigued by the different approaches I've seen in the wild, and I have a rich editorial calendar of content to share with you about these findings (when I get the time).

What an Agent Score Can Tell You

Sat, 18 Apr 2026 16:00:00 -0500

If you're a documentation team trying to figure out what "agent-ready" means for your site, you've probably noticed that several different organizations are now offering to score you. Cloudflare just launched isitagentready.com. Fern has an Agent Score directory powered by my Agent-Friendly Documentation Spec and its CLI tool, which runs 22 checks against any docs site. There will probably be more by the time you read this.

Each of these tools produces a number. Numbers are comforting. They give you something to put in a slide deck and something to improve quarter over quarter. But if you're making decisions about where to spend your team's time based on these numbers, you need to understand what each score actually measures. They're not measuring the same thing, and the gaps can be dramatic.

Designing an Agent Reading Test

Mon, 06 Apr 2026 11:00:00 -0500

With my work on the Agent-Friendly Documentation Spec and the companion AFDocs tool, I've been talking with a lot of people about what all this means. I've been writing about the different failure modes in a way that might make sense to a web developer, or a documentarian with a strong technical understanding of how websites work, but there are a lot of pieces involved. It's not just about the website behavior; it's also about agent platform limitations. The intersection of those things is where the failures originate. And because all this requires a strong technical understanding of both parts, I realize it might be harder for people who are not super involved in this space to understand the impact.

Measure Agent Web Traffic Redux

Sat, 04 Apr 2026 20:00:00 -0500

Almost exactly a month ago, I wrote an article about poking around in my server logs attempting to measure agent web traffic: How to Measure Agent Web Traffic. I was trying to give our documentation platform team information to help them identify agents in our web traffic. My conclusion at the time was this:

So when it comes to user agents, I don't think we can reliably correlate a specific user agent string with a specific agent platform. But there are patterns we can look for that are probably generically "agents" - and that may have to be good enough for now.

How to Evaluate a Platform-Written Spec

Sat, 28 Mar 2026 15:00:00 -0500

Recently, Vercel published an Agent Readability Spec (Timothy Jordan, last updated March 23, no publication date), a scoring system for how well websites serve AI agents. It covers llms.txt, markdown mirrors, content negotiation, structured data, robots.txt, and a checklist of other recommendations. Four days later, they followed up with "Make Your Documentation Readable by AI Agents" (Rich Haines; March 27), a companion implementation guide that references the spec article and adds its own recommendations. If you're a documentation team trying to make your site work better for coding agents, either article looks like a useful starting point.

When a Feature Request Becomes a Research Project

Sat, 21 Mar 2026 19:00:00 -0500

Someone made a feature request on the skill-validator tool: support for the evals/ directory introduced by Anthropic's recent updates to its skill-creator skill. This is the third or fourth Claude Code-specific request I've gotten regarding skill-validator development, and the tool is quickly accumulating custom flags to enable support for Claude Code-specific functionality while trying to preserve general spec adherence for folks developing cross-platform skills.

I've told people on more than one Issue or PR that I intend to look into cross-platform support for Skill file loading, so I can have more concrete data than "I hope citing the spec here means it will be portable across platforms." After yesterday's article about the Agent Skill governance issues, and the lack of support and guidance that community members are getting from spec maintainers, I decided it was time to stop saying "it's on my todo list" and put together some resources to help other people dig into it.

Why a Platform Shouldn't Own an Open Spec

Fri, 20 Mar 2026 15:00:00 -0500

The Agent Skills spec claims to be "A simple, open format for giving agents new capabilities and expertise." But the creator of Agent Skills, Anthropic, is also the company behind the spec's maintenance. And when a platform both owns a spec and competes with the platforms implementing it, the incentives stop being aligned. Let me tell you a tale of how Anthropic has quietly broken Agent Skill interoperability with every other platform, encoded it as ground truth in a spec that claims to be open, and what that means for anyone publishing or consuming skills with portability in mind.

Is Your llms.txt Already Stale?

Sun, 15 Mar 2026 23:00:00 -0500

You shipped your llms.txt. You linked to your docs pages. Maybe you even set up progressive disclosure with per-product files. You're done, right?

Probably not. An llms.txt that was accurate when you launched it can silently drift out of sync with your actual documentation. New pages get added to the site but not to llms.txt. Old pages get removed or reorganized. The file becomes a stale index that points agents toward an incomplete picture of your docs, or worse, toward pages that no longer exist.

Agent Skill Mega Repo Woes

Fri, 13 Mar 2026 07:00:00 -0500

A month ago, I published an ecosystem-scale analysis of 673 Agent Skills from 41 repositories. I looked at Anthropic's own skills, company-published collections from Microsoft and Stripe, community collections, and individual repos. The takeaway was that the ecosystem has real quality problems: 22% of skills fail structural validation, over half of all tokens are wasted on non-standard files, and broken links are everywhere.

Today, I saw a post on LinkedIn about antigravity-awesome-skills, a single mega repo that claims over 1,200 Agent Skills and has 23.7k GitHub stars. That star count makes it one of the most popular skill sources in the ecosystem. People are presumably cloning this repo and loading these skills into their agents.

Vibes are Out, Data is In

Thu, 12 Mar 2026 07:30:00 -0500

There's a growing body of advice about how to write documentation for AI. Some of it is grounded in real observations. A lot of it is grounded in vibes: intuitions extrapolated from LLM research that was never designed to test documentation, recommendations passed around conference talks and blog posts until they calcify into "best practices," and vendor guidance that may or may not reflect how agents actually behave in the wild. People aren't wrong to form theories. But we're skipping the step where we test those theories in the context that actually matters, and then building an industry playbook on top of the untested assumptions.

How to Measure Agent Web Traffic

Thu, 05 Mar 2026 21:00:00 -0500

After sharing my recent findings about how agents use docs with the folks at work, one of the first questions (from multiple people!) was: how do we measure the impact of these problems? Writ large, how do we know things like:

How many agents are trying to access our docs and running into issues?
How many docs pages are adversely impacted by these issues?
How can we tie agent access issues back to developer/product outcomes?

These are questions with many facets, and we're still talking about what the answers might look like. But the first two problems seemed immediately tractable, so we can start digging into those things right away.

An Agent is More Than Its Brain

Mon, 02 Mar 2026 20:00:00 -0500

Apparently I live in the "agent" space now, so let's talk about what's inside an agent! I've been seeing conversations, and learning more myself, and have realized that maybe a lot of people haven't stopped to think about what the component parts of an agent actually are. For the purposes of this article, I'm talking specifically about coding agents, but a lot of the bits and bobs are generalizable. And here's why this matters: the model ("Opus 4.6," "GPT-4o") is just one piece. The same model can behave very differently depending on everything else around it.

Make Your Hugo Site Agent-Friendly

Sun, 01 Mar 2026 08:00:00 -0500

With all the writing I've been doing lately around agent-friendly docs, I decided that all new websites I set up will be agent-friendly from the beginning. I designed my two newest websites, the Agent-Friendly Documentation Spec and aeshift, to be agent-friendly from day one. Particularly with the spec, I thought that people might want to point agents at the spec, and I wanted to enable that. I pointed an agent at the spec (as local markdown on my filesystem) and built the sites to hit all the agent accessibility criteria. It didn't add that much extra time to setting up each site, and it got even faster after I had a reference implementation I could point the agent at.

Case Study - 'upgrade-stripe' Agent Skill

Fri, 27 Feb 2026 08:00:00 -0500

When I did the research for my recent Agent Skill research paper, I took two approaches to trying to understand the impact of Skills in developer workflows: take a wide, zoomed-out look at Skills across different segments, verticals, and use cases, and take a closer look at a subset of those Skills where I had interesting theories I wanted to probe more closely. One of the skills I took a closer look at was the upgrade-stripe Skill. I had ideas about what I might find, and thought it might have implications for how other companies with multiple programming language SDKs should approach Skill development. What I found was not what I expected, but was arguably more interesting.

LLMs vs. Agents as Docs Consumers

Thu, 26 Feb 2026 08:00:00 -0500

I've been writing a lot about agents and docs lately, and one thing I keep bumping into in conversations is confusion about what "AI-friendly docs" actually means. Someone says their leadership has mandated that docs need to be optimized for AI, and when I ask what that means in practice, the answer is usually some variation of "I don't know, they just said AI." And honestly? I get it. "AI" is doing a lot of heavy lifting as a term right now, and it's papering over a distinction that really matters if you're trying to figure out what to actually do with your documentation.

Agent Web Fetch Spelunking

Thu, 19 Feb 2026 21:00:00 -0500

After posting about Agent-Friendly Docs, I got some very good questions that I wanted to dig into. I love the documentation community - such thinkers! So today, I present my spelunking in the agent Web Fetch tool. My goal was to figure out if I could provide a quick answer to whether agent platforms document the truncation limits of their various web fetch implementations, so documentarians can try to wrap our heads around what makes "agent-friendly docs." First criteria: agent must be able to actually get the docs content.

Agent-Friendly Docs

Wed, 18 Feb 2026 21:00:00 -0500

Over the weekend, I spent about 10 hours working with Claude to manually validate 578 coding patterns across a range of languages and coding tasks for my Agent Skill Report. We had to deep dive into standard library docs, package-specific documentation, large enterprise docs, and personal blogs. My goal was to ensure that every API and configuration pattern I was using in a behavioral evaluation experiment reflected the current guidance from official sources. My learnings about agent docs access patterns were secondary to this project, but the results are a rich treasure trove of data that may inform how I use agents with docs in the future. As a technical writer who has worked on SaaS and developer docs for the last decade, honestly, my findings made me a little sad.

Agent Skill Analysis

Fri, 13 Feb 2026 10:00:00 -0500

Claude introduced Agent Skills in fall 2025, and they have quickly become the next hot thing in AI-assisted development. Agent Skills give your AI buddy just-in-time context to help it succeed with tasks that require specific domain expertise or resources. Anthropic released an Agent Skills specification, and other agentic platforms have been rushing to add support for it. In turn, individual developers are making their own Agent Skills, and companies are under pressure to provide official Agent Skills for customers to use in agentic development workflows.