AI Agents Now Outnumber Human Users on Our Skill Directory
We run a directory of 590+ agent skills. Every install — whether it's a human clicking a download button, a developer piping curl into bash, or an AI agent fetching an install manifest — goes through the same endpoints. Thirty days of that data just told us something we didn't expect: agents are now our dominant user class.
The numbers
Last 30 days, 4,242 install events across three channels:
| Channel | Installs | Share |
|---|---|---|
| AI agent direct (install manifests fetched by agents) | 3,689 | 87% |
| Human via curl one-liner | 320 | 7.5% |
| Human via web download button | 183 | 4.3% |
Read that again. For every human who clicks a download button, twenty agents fetch and install skills on their own.
How we can tell agents from humans
Three signals, triangulated:
- User-agent: agent installs come from CLI and SDK user agents, not browsers.
- Referrer: humans carry page referrers; agents don't.
- Traffic shape: agents skip the website entirely — no pageviews, just manifest fetches. A country shows 0 visitors but 12 installs? That's agents working behind a proxy in that region.
We split our stats into human and agent columns after noticing countries with "impossible" data — installs with zero visits. It wasn't a bug. It was a new user class.
Why this happened: llms.txt and machine-readable install manifests
Two decisions we made earlier this year primed the pump:
We published an llms.txt file — the emerging convention for telling LLMs what your site offers. Agents that research before acting read it.
Every skill page exposes a machine-readable install manifest at /install/{slug} — plain markdown, no JS required, with three install paths (curl one-liner, AI-assisted prompt, manual download). An agent can read it, decide, and execute without any human in the loop.
The result: when a user tells their agent "find me a good PDF skill and install it," the agent searches, finds our directory, reads the manifest, runs the install. The human never visits our website. They don't need to.
What this means for skill distribution
The website is no longer the product — the install endpoint is. Our pageviews are modest; our install counts are not. If you're building a skill directory, a tool registry, or any agent-adjacent resource, optimize for the machine reader:
- Publish an llms.txt. Agents can't install what they can't discover.
- Make install manifests plain and parseable. No JavaScript-gated content, no auth walls, no cookie banners for machines.
- Version your install endpoints. Agents cache; stale manifests break installs silently.
- Instrument the agent path separately. If your analytics only track pageviews, you're blind to your biggest user class.
The uncomfortable part
87% agent installs also means our numbers can be gamed. Nothing prevents a malicious actor from pointing a botnet of agents at a target skill's install manifest to inflate its stats. We've kept human and agent columns separate precisely so that "most installed" rankings can't be captured by whatever bot is loudest. If you run a directory, do the same from day one — retrofitting the split after the fact is painful.
What's next
The agent share is still climbing. We're now seeing multi-skill installs — agents fetching five manifests in a session after a user asks for "a complete writing setup." Our scenario wizard (pick your role, get five matched skills, one prompt to install them all) was built for humans; agents use it too.
The browser isn't dying. But for skill distribution, it's already the minority report.
We evaluate every skill on six dimensions before listing — see the methodology. Install data reflects 30 days ending Oct 8, 2026.
Top comments (0)