Skip to main content
An agent in Agent Registry is a registry record for an AI agent, linked to its sessions, usage, and runs. The registry never runs the agent. It records what the agent did, wherever it ran, so you can inspect one agent or compare the whole fleet.

​
Agents and harnesses

Agent Registry records two kinds of AI worker: The difference is who drives the work: a person at a harness, or the agent’s own code.

​
How runs reach Agent Registry

The registry executes neither harnesses nor agents. A run reaches the registry as a record, in one of these ways:
  • Session upload: a run’s record and transcript are uploaded after it happens. The desktop app does this for Claude Code and Codex sessions on your computer.
  • Traces: the trace plugin for Claude Code, the atlanai SDK, or any OpenTelemetry (OTLP) exporter sends the run’s execution evidence.
A recorded session is final. It cannot be edited or deleted afterwards.

​
Agent page

An agent page has the tabs Overview, Sessions, Usage, and Relationships. Releases also appears for agents that support releases.
  • Sessions lists the recorded runs for the agent.
  • Usage shows runs, models, tools, cache use, and recent runs for the selected period. Select a recent run to open its trace.
  • Relationships shows the agent’s links to other registry objects.
To compare agents instead of inspecting one, use Agent usage. It summarizes usage, cost, efficiency, and recorded errors across every agent with recorded runs.

​
Example

A support team builds an agent with LangGraph and sends its traces through the atlanai SDK. In Agent usage, the team sees that the agent’s errors rose this week. They open the agent’s page, select Usage, and open a failed run from Recent runs. The run’s trace shows the step where a tool call failed.

​
What agent records do not prove

  • An upload does not run the agent. Creating a session records a run; it does not start one.
  • An accepted upload does not prove completion. Compare the session’s outcome, timestamps, and final events with your runtime when it matters.
  • Missing data is unknown, not zero. An agent with no usage has no attributable runs, which is not the same as no activity.

​
See also