Run your first task

Start with a complete example, or run a task and agent you already have.

Run your first agent on a single task. The complete example supplies the task, agent profile, environment and assessment, along with a recorded result you can inspect before running it yourself.

The tutorial uses code review as an example. You can use the same task, profile and assessment structure for other kinds of work.

Use existing inputs

Already have a task and compatible agent? Follow the path below. One execution does not establish that the agent is good at your broader workload.

Before you start

Create a Synth account, install the CLI and sign in.

Choose a task and agent

Find tasks and agent profiles available to your account:

synth task list
synth agent list

Choose a small task with a verifier and an actor profile appropriate for its work. If you are starting from scratch, use the supplied example, or create your own task and agent profile. Replace the placeholders below with your resource references.

Replace <task-ref> with the selected task reference and <actor-ref> with the selected agent:// reference:

synth task show <task-ref>
synth agent show <actor-ref>

Keep the exact version references returned by these lookups for the launch below. Read the task’s instructions and required output, and check the profile’s model and execution recipe. Listing or resolving a profile does not prove that it is compatible with every task or that its model is available to your account.

Inspect, then launch one trial

synth run start --task <task-ref> --actor <actor-ref> --dry-run

Check that the request selects your exact task and profile. The dry run does not execute the agent or prove the task will succeed. Remove --dry-run when you intend to start one real managed trial:

synth run start --task <task-ref> --actor <actor-ref>

The default is one trial with the task’s verifier. This submission uses compute. Submit once and save the returned run ID. If the response is uncertain, inspect the existing run before retrying.

Read the result

synth run show <run-id>
synth run events <run-id> --follow --timeout 10m

If the watch ends while the run is still active, inspect synth run show <run-id> again; ending the local watch does not cancel the run. After it reaches a terminal state:

synth run result <run-id>
synth run trials <run-id>
synth run export <run-id> --output ./<run-id>-evidence

Export after the run is terminal, into a new directory, replacing both occurrences of <run-id> with the same ID. Check the manifest and inspect the agent’s work and verifier result in the retained artifacts. A nonzero process exit or missing output is not the same as an assessed incorrect answer. If execution fails or stops progressing, use the troubleshooting guide before submitting another run.

You have completed the path when you can identify the selected task and agent, explain the terminal outcome and locate the evidence. A poor score is still a result to investigate, not a reason to claim execution failed.

Take the next useful step

• Understand assessment, including its limits.
• Bring your own task.
• Repeat and compare before choosing an improvement.
• Train and test a checkpoint once the task and scoring are useful.