An engineer at Linear shares his experience building automations that use the Linear Agent to auto-fix bugs in triage. Two workflows proved effective: cleaning up dead feature flags (near 100% success rate) and patching failing background tasks (about one-third success rate). Key lessons include narrowing the scope of tasks, adding gates that force the agent to prove it understands a problem before attempting a fix, and encoding team conventions in both Linear Skills/Automations and the codebase itself. The author also describes a task-splitting workflow where large ambiguous tasks are broken into well-defined pieces for the agent, while the engineer handles complex parts in parallel. The session-based, multiplayer nature of Linear Agent — where all context stays attached to the issue — is highlighted as a differentiator from editor-centric coding tools.
