Skip to content
Merulan

Build and launch work

Save an Agent, launch its saved version, and check the Run, review boundary, and retained result.

Use this guide when you need to build and launch work in Agents & Workflows. Start with one saved Agent, then read the Run it creates.

Before you start

A builder needs an Agent they may edit and run in the selected Workspace. Have a bounded prompt and a result you can check. Review the Agent's tool scope, credentials, and approval needs. The Console path uses Save & Run to execute one saved Agent in a one-block Workflow. Use the start contract for the exact request, durable identity, and supported Surface.

1. Build the definition

Build the Agent at /workflows/agents/[agentId] in the Console. Save its instructions and allowed tools. Read the accepted version before launch. If the job needs several blocks, compose them in the Workflow template editor at /workflows/definitions/[workflowId]. Publish the draft after its graph, accepted Agents, launch settings, and review rules pass validation. Publication freezes the version. It does not execute the work.

2. Launch saved work

Launch the Agent from its Save & Run panel with the bounded prompt. The panel saves before starting the Run. When the definition is read-only, Run uses the saved configuration. Check that the panel shows the retained Run, Agent version, and fingerprint. For a multi-block Workflow, use its configured launch path after publication. That path checks the published launch rules before creating a Run.

3. Read the Run

Read the Run shown by launch and follow its current state. Open the Run detail for the block result and error, if any. When a call waits for approval, use Reviews and repair to decide the exact request. Read the Run again after the decision. Approval clears one gate. It does not waive the call's permission or credential checks.

Checkpoints

Build and launch has three checkpoints: an accepted definition, a retained Run identity, and readable execution state. Match the recorded Agent version to the saved work you intended to run. A start response proves admission, not completion. Save & Run starts a separate Run for each fresh submission. Do not select it again to refresh an existing Run's state.

Success

Launch succeeds when the retained Run can be read with the expected saved version. The work succeeds when the Run reaches completed and its result meets the prompt's goal. A failed or cancelled result is still useful evidence, but is not a successful outcome. Keep the Run for comparison when a later definition needs a separate attempt.

When it fails

Build and launch failures can occur before admission or during execution. Read the reported error and retained state to tell those cases apart.

AGENT_NOT_RUNNABLE

The saved Agent lacks policy enrollment. Recreate it through the Console, as the refusal directs, then review its definition and grants. A copied configuration does not substitute for enrollment.

RUN_START_IN_FLIGHT

The same start request is still being admitted. Read back the existing attempt before asking for different work. A fresh Save & Run submission creates a new request, so repeated submissions can start separate Runs.

RUN_START_FAILED

The retained start request already failed. Repair the reported cause before making a separate attempt. If a Run was created and then failed, inspect its block error through Run readback instead.

Recovery

Build and launch recovery preserves the earlier evidence. Correct a draft refusal before publishing again. For a Run that needs review, retry, or resume, follow Reviews and repair before starting another Run. A lost browser view does not stop execution. Reopen the retained Run and read its state. Stopping observation only stops the browser's polling.

Next steps