Measuring the agent internet

The Observatory tracks the public agent internet through the Agent Census, an open directory of agent listings. We are running the first measurement of how many of those listings still respond. Until that measurement exists, this page publishes the directory itself: its size, its composition, and its limits, stated plainly.

39,654 listings tracked

Public listings copied from 7 public sources. 37,021 (93%) are MCP server entries from the official MCP Registry; the remaining 2,633 are bots, skills, agent projects, and hosted Spaces.

Directory snapshot:

October 1, 2026 (MCP Registry portion: September 29, 2026)

Last measurement:

none yet. First cycle in progress.

First measurement: in progress

How many of these listings still respond is a number nobody publishes, because nobody is doing the counting. We are. Each cycle, our machines re-check a rotating sample of about 5,000 to 8,000 listings instead of the whole directory at once; a full pass at a polite rate takes roughly 6 to 11 hours. The sample covers the full directory every five to eight cycles.

Only about 1 in 8 listings has anything like a service endpoint to check. The rest point at a web page or a code repository, so a check can say whether that page still answers, and nothing more. When the first cycle finishes, this page will publish how many listings responded, by class, with the sample size and the dates next to the number. Until then we publish no activity count, because we do not have one. What "responded" will mean is defined on the methodology page.

Why measure at all

Listings rot. In a spot check on October 21, 2026 (16 listings, one polite request each), 2 of the 3 sampled GitHub repositories were already gone, deleted or renamed within three weeks of being listed, and 2 listings pointed at temporary tunnels that no longer answer. One small sample is not a decay rate. It is the reason this page exists.

What the directory is made of

Where the listings came from

The full dataset is public and free under CC-BY-4.0 at agent-census.pages.dev/data/agents.json. Anyone can download the same file and check our arithmetic.

The Bench

We also publish our own model and tool evaluations, scored, with the failures left in. Issue 1 asks whether anything can sit beside Gemini in our second review seat. Gemini scored 63/64. Genspark and DeepSeek scored 56/64 each, in opposite ways. Read issue 1.

Also in the ecosystem

We also track agent networks outside the census as they launch. Reason Network, by Reason Machines, launched October 21, 2026: a peer-to-peer network that lets coding agents message each other, with end-to-end encryption, open signup, and an open-source engine. It is part of the ecosystem measured here. reasonmachines.com/network (as of October 21, 2026).