Build an AI agent that gets delegated a Linear issue, works an Agent Session, and reports structured progress back to the user

Documented steps

1. Authorize with actor=app and relevant agent scopes (app:assignable for delegation/project membership, app:mentionable for cross-issue/document mentions) so the app can be installed and delegated to.
2. Listen for the AgentSessionEvent webhook: a `created` action fires when a user mentions or delegates an issue to the agent; a `prompted` action fires when a user sends a new message into an existing session.
3. Within 10 seconds of a `created` event, emit a `thought` AgentActivity to avoid being marked unresponsive — use the event's promptContext field to build context from relevant issue details, comments, and guidance.
4. If the delegated issue isn't already in a started/completed/canceled status, query workflow states filtered by type: { eq: "started" }, pick the one with the lowest position, and move the issue into it; also self-assign as delegate if no delegate is set yet.
5. As work progresses, create further AgentActivity entries via the agentActivityCreate mutation (or SDK's createAgentActivity) using the appropriate type: thought (reasoning), action (tool calls with a result), elicitation (need user input), response (finished), or error (failure).
6. Reconstruct conversation history by querying Agent Activities (immutable snapshots) rather than issue comments, since comments can be edited by users after the fact.
7. Set externalUrls via the agentSessionUpdate mutation (e.g. a dashboard or PR URL) so users have visibility into out-of-band work and the session doesn't look stalled; let Linear manage session state (pending, active, error, awaitingInput, complete, stale) from your activity emissions.
8. Official docs: https://linear.app/developers/agents plus /agent-interaction and /agent-best-practices