AgentariumRESEARCH OBSERVATORY
Loading catalog

AGENTARIUM / RESEARCH OBSERVATORY

Observing the
agent ecosystem.

A source-linked catalog of AI agents and the work they claim to perform. Explore the systems, examine their tasks, and follow the evidence.

Explore the observatory Browse the dataset
–AGENTS
–TASKS
–QUOTED TASK CLAIMS

CLAIMS QUOTED FROM PUBLIC SOURCES · NOT MEASURED CAPABILITIES

OBSERVATION FIELD / 001
Each point is one agent's quoted claim to a task.Inspect a point →

An agent with several documented tasks appears once in each task's column.

YASAMAN ANSARI / AI AGENT RESEARCHEXPLORE THE CATALOG UPDATED DAILY

01 / AGENT HABITAT

An ecosystem, in view.

Every creature is an agent in the catalog. Form and colour show a documented task; select one to read its record and the quotes behind it.

0 IN VIEW
COLOUR + SPECIES = DOCUMENTED TASKSIZE = NUMBER OF DOCUMENTED TASKS
A random selection from the catalog, drawn again on every visit.View tabular records →

02 / FINDINGS · TASKS

The distribution
of agent work.

FIGURE 01

Tasks, clearly defined.

Each task comes from one shared vocabulary, and an agent is linked to a task only when its own text says it performs that task, with the quote kept as evidence.

This view counts agents in the catalog. It describes what the catalog covers, rather than estimating the size of the market.

Agents by documented taskCOUNT

An agent with several documented tasks is counted under each. Select a bar to explore that species in the habitat.

03 / FINDINGS · TIME & OWNERSHIP

Patterns across the catalog.

When the catalog first recorded each agent, and who publishes it, give two more views of the same records.

FIGURE 02 / TIME

When the catalog found them

Agents by the month the catalog first recorded them.

The first-seen date is when the daily collection first recorded an agent, not when the agent was released. Colour shows each agent's first documented task.

FIGURE 03 / PUBLISHER

Who publishes the agents?

Share of agents with each task, by where the record comes from.

GitHub records are split by whether the repository belongs to an organization or a personal account. Publishing an agent does not establish who uses it.

04 / DATASET

Inspect the underlying records.

Search the catalog, compare documented tasks, and inspect the record behind each agent.

Agent catalog. Select an agent name to inspect its record and evidence.
SOURCEDOCUMENTED TASKSPUBLISHER
Every task links to a quote from its source

05 / METHOD & EVIDENCE

A catalog built
around its sources.

Clear definitions, consistent identifiers, and evidence attached to each documented task claim.

01 / COLLECT

Collect public evidence.

Collect public agent repositories and preserve the source text used to describe them.

02 / CLASSIFY

Classify documented tasks.

Assign tasks from a shared vocabulary. Keep a source URL and a supporting quote for every task claim.

03 / CONNECT

Connect the records.

Link agents and tasks with stable IDs. Validate the CSV tables and publish a version researchers can identify.

THREE TABLES. ONE CATALOG.
agents.csv↔agent_tasks.csv↔tasks.csv

A documented claim is not a measured capability. Source coverage shapes what the catalog can show.

0000/Overview

DATA DICTIONARY

Three tables. Clear connections.

agents.csv

One row per agent. A permanent agent_id, display name, description, homepage, source URL, and first-seen date.

tasks.csv

A shared task vocabulary. Each task_id has a name and definition, so equivalent tasks use the same label.

agent_tasks.csv

One agent–task connection, identified by agent_id + task_id, with a source URL and a verbatim evidence quote.

Download the full tables: agents.csv · tasks.csv · agent_tasks.csv. A documented claim records what a project says it does. It is not a measured capability.