Agents

OverviewCopy Anchor Link

An agent is a purpose-built version of Inspect. It combines a system prompt, model, repository access, capabilities, and optional skills. Its name and handle let teammates invoke the same specialized behavior from the web app or Slack.

Create and manage agents in Customize → Agents.

Reference a custom agent from SlackCopy Anchor Link

After a custom agent has been added to Slack, mention its Slack handle and write the request in the same message—for example, @project-manager summarize Project Atlas’s current status, blockers, and next milestone. The agent starts a session with its configured prompt, model, skills, and access, then replies in a thread.

If the agent uses Inspect’s shared Slack app, invite @Inspect to the channel first, then mention the custom agent’s handle. If its handle does not appear in Slack, open the agent in Customize and add or install it to Slack.

Workspace administrators can create a public Slack channel while selecting the channels where an agent automatically replies. Enter a new channel name under Slack channels, then select Create. Inspect adds the new channel to the agent's selection.

Use a custom agent in a child sessionCopy Anchor Link

An Inspect session can spawn a child session using a specific custom agent. Ask the parent session to use the agent by name; Inspect finds the matching agent and starts the child with its exact agent ID. The child runs independently in its own workspace with that agent’s configuration and reports its result back to the parent.

Set boundariesCopy Anchor Link

Grant only the repositories and tools the agent needs. Describe what it should deliver, when it should ask for input, and any standards that must always apply.

Improve over timeCopy Anchor Link

Agent prompts are versioned. Compare behavior across real sessions, refine the prompt, and restore an earlier version if a change reduces quality.

Real-life examplesCopy Anchor Link

Design QA agent: knows the design system, checks responsive states, and reports evidence without editing backend code.

Data investigator: can query approved analytics tools, understands company metrics, and returns a concise decision-ready report.

Migration agent: follows a codemod playbook, works only in approved repositories, and applies the same verification standard to every batch.

Combine agents with automationsCopy Anchor Link

An automation starts an agent when its trigger fires. In the automation editor, choose the agent and write a short prompt for that trigger. Then select where the result should go. Each run follows the agent’s instructions and uses its access settings.

A GitHub automation can send UI changes to a Design QA agent. The agent checks the changed screens and posts screenshots with a ranked list of problems.

A Datadog automation can send monitor transitions to a Production Triage agent. The agent follows the team’s playbook, reviews the relevant telemetry and recent deploys, then posts its assessment to the service channel.