Agent

An agent is an AI agent you have connected to Prefactor — a conversational assistant, an automated workflow, a retrieval pipeline, or any other AI system whose behaviour you want to govern.

Every agent has a stable identity in the platform. That identity is the anchor for everything Prefactor tracks on its behalf: the runs it produces, the versions it has, the activity schema that describes its span types and payloads, the risk profile that governs how the data risk of its runs is classified, and the agent deployments that deploy specific versions to specific environments. Without that anchor, spans and instances would be floating data with no way to compare behaviour over time or tie a compliance review to a specific agent.

An agent belongs to your account, not to any single environment. You register an agent once; it then shows up in every environment where it produces runs. Each run is recorded as an instance under that agent, against the version specified when the run was started.

Registering an agent in the platform creates the named entry. Your integration identifies itself to Prefactor at runtime through a deployment-scoped API token, which carries the agent identity automatically.

Versions

Each agent accumulates a history of versions. Prefactor creates a version on demand: when an instance arrives reporting a combination of metadata and schema not seen before for that agent, a new version is created automatically — the same mechanism that creates new activity schema versions. The external identifier reported with the instance (a semver tag, a commit SHA, or any string your deployment uses) becomes the version’s name.

Versions exist for traceability. Every instance is tied to the version it ran under, and every version is paired with the activity schema version registered alongside it, so you can always trace a run back to the exact code and schema that produced it.

When you deploy a version to an environment, Prefactor records that as an agent deployment. The Versions tab shows which environments each version is currently deployed in.

In the Admin UI

• Agents page — the full list of agents in your account, with a 24-hour activity summary.
• Agent page — identity, the 24-hour activity timeline, deployment snapshot across environments, and how the tabs fit together.
• Agent page › Overview tab — instance counts by state and recent spans with drill-through to instances.
• Agent page › Versions tab — the full version history for an agent, with schema version associations and deployment status.

Related concepts

• Account — the boundary that owns agents, environments, and team access.
• Activity schema — describes each agent’s span types and payloads; each version is paired with an activity schema version.
• Agent deployment — deploys a specific version to an environment as the current release.
• API token — deployment tokens carry agent identity at runtime.
• Instance — a single run of an agent; every run is attributed to one agent and one version.
• Environment — controls which activity is visible, not which agents exist.
• Risk profile — risk profiles are assigned to agents and classify the data risk of their runs.