Guide
Your first task, step by step
A practical walkthrough of the console: launching a task, steering it with approvals, and reviewing what it produced. It takes about five minutes on the mock runtime.
Run your first task#
- 01
Pick a workspace and model
On Tasks, choose the repository the agent will work in and the model profile that powers it. Each workspace runs in its own sandbox with no network access.
- 02
Describe the goal
Say what you want and how you'll know it worked. Click one of the example chips to start quickly. Open Budget & success criteria to set limits and add explicit acceptance checks, one per line.
- 03
Choose a mode
Supervised asks before every file write and protected action, which is best while you build trust. Bounded autonomous runs routine actions on its own but still stops for network, installs and destructive operations.
- 04
Launch and watch the timeline
Press Launch task (or
⌘ ↵). The timeline shows every state change, tool call, test run and agent note as it happens. The plan on the right tracks progress step by step, and the meters at the top show budget use. - 05
Approve or reject protected actions
When the agent needs permission, an amber Approval required panel shows the exact parameters and risk. Approving is bound to those parameters. If the agent revises its patch, you'll be asked again.
- 06
Review the result
When the task succeeds, read the Final report and its evidence, inspect the Diff tab, and download patches or logs from Artifacts. Nothing is pushed for you: apply the patch to a branch yourself.
Writing good goals#
The best goals name the target, the expected behaviour, and a check that proves it. Vague goals produce vague plans.
Fix the bug.
Fix the invoice total being off by one cent when unit prices have 3 decimals. test_three_decimal_unit_price_rounds_half_up should pass.
Add rate limiting.
Add rate limiting to POST /api/login: max 5 attempts per minute per IP, return 429 when exceeded. Keep existing login tests green.
Clean up the code.
Refactor internal/sink/csv.go to stream rows instead of buffering the whole file. No behaviour change; go test ./... must pass.
When a task stops#
- Budget reached. The task ends as Failed · partial result with whatever diff it has. If it's marked recoverable, press Resume: it continues from persisted state and extends the limits it used up.
- You rejected an action. The task ends as blocked. Refine the goal or constraints and start a new task.
- You changed your mind. Press Cancel at any time. The worker is terminated and nothing leaves the sandbox.
Try it
0 before launching to watch budget enforcement stop a task.Tips & shortcuts#
⌘ ↵/Ctrl ↵: launch the task from the goal box.- Tick show policy checks on the timeline to see every allow / approval / deny decision.
- Tasks waiting for you float to the top of the task log with an amber marker.
- The mock runtime saves tasks in your browser; a reload recovers in-flight tasks as resumable. In API mode tasks live in the control plane.
- Switch the interface language from the sidebar. Task data and logs stay in their original language.