Starting runs

A run executes a prompt for an agent. You can start one yourself, call Jazz from a script, or arrange for a schedule or external event to start it later.

Commands and workflows

CLI: jazz agent chat <name> opens a conversation. Each message starts a run for that agent.
Headless command: jazz run executes one prompt from an argument or stdin. Use it in scripts, CI, or for a one-off task.
Workflow: jazz workflow run <name> executes a prompt saved in WORKFLOW.md, along with its agent, permissions, and limits.

Schedules

Scheduled workflow: jazz workflow schedule <name> runs a workflow on its cron schedule. Use it for recurring jobs such as a daily briefing.
Loop: /loop 10m <prompt> repeats a prompt until its purpose is met or you stop it. Use it to check a deployment or watch for a change.

Events and deferred work

Webhook: an authenticated HTTP request starts an agent with a fixed prompt template and the caller’s payload. Use it for alerts and application events.
Peer request: another configured Jazz agent sends a question to your agent, subject to the permissions you granted that peer.
Wake trigger: the agent schedules a later run in the same conversation. Use it to follow up on work already in progress.
Background jobs: the agent starts independent shell jobs, then resumes its conversation with their results when they finish.

A reminder sends a note without starting an agent run.

Permissions and results

A run started without someone at the terminal uses its configured approval policy. Actions outside that policy are declined or parked for later approval, depending on the surface. With jazz run, use --park to save a run that needs approval. See Unattended runs.

Set budgets before scheduling recurring work. Configure notifications to receive results, failures, and approval requests while you are away. A scheduled workflow’s deliver: field selects where its results are sent.