Your first task

Create a task from the dashboard or the CLI, follow it through verification, and read its completion report.

Create it

Dashboard: Tasks → New task. Choose the project, give a title and describe the change as you would to a colleague: what, where, and how you will know it works.

CLI:

Terminal

agentctl task create --project prj_123 --title "Validate email on sign-up" \
  --prompt "Reject sign-ups with invalid email addresses in the API and show the error in the form. Add tests."

What happens

Scheduling. The control plane offers the task to a worker that has the project, a compatible agent and free capacity. If none qualifies, the task stays QUEUED and says what each worker is missing.
Preparing. The worker checks out a task branch (ao/…), gathers skills, MCP servers and plugin instructions, and writes the agent-neutral instructions.
Running. The agent works in the project folder. Progress, tool calls and checkpoints stream to the task's timeline.
Verifying. The worker runs the project's checks — detected from package.json, Composer, Python, Go or Rust projects, or configured in the policy. Failures go back to the agent.
Git. With the default policy COMMIT, the verified change is committed on the task branch.

Read the report

The completion report lists what changed, which checks ran and their results, the Git result, and a summary. Verified facts and the agent's own claims are kept apart, so you can see what was proven.

When it does not complete

State	Meaning	Do
`WAITING_FOR_LIMIT`	The agent hit a usage limit	Wait, or add add-on models
`WAITING_FOR_INPUT`	The agent asked a question	Answer in the dashboard or `agentctl task input <id> --text "…"`
`RECOVERY_REQUIRED`	Retries or remediation are exhausted	Read Recovery on the task, fix the cause, Retry