Troubleshooting

Start with the message or state shown in your world. For the duplicate-invoice example, use the same ENG-2, order 7001, case 00001001, and #erp-alerts references when checking records and scores.

​

Worldsmith asks questions instead of building

Answer the displayed questions and choose Continue. Name the workflow, services, and success criteria you care about, and tell Worldsmith which other details it should choose. Use a complete brief if you want a starting point.

​

Launch world is disabled

Wait for generation and checking to finish. The draft must include at least one service and have no compile problems before Launch world is available.

• A few things need fixing: correct the listed records or scorers, or choose Fix with Worldsmith.
• Couldn’t check this world: read the error and choose Check again.
• World prepared with notes: expand and read them. Warnings do not block launch.

If you are editing a saved world, the launch action is named Launch version or Relaunch. See Refine the next version.

​

Couldn’t launch this world

Read the displayed error, then use the button shown: If the error names twin capacity, check that the organization has room for the included services. A four-service launch needs capacity for four twins. A failed launch can leave some twins running; inspect them before relaunching. For automation-specific retry behavior, see Errors and retries.

​

The world shows Needs attention

Open Twins and inspect each service’s status and error. Needs attention can mean an unhealthy twin or a mix of running and stopped twins. Confirm that every service the task needs is running before using that world. If you need a fresh copy of the saved configuration, choose Relaunch. That creates a new version with its own twins; it does not repair the old version’s records in place.

​

The world is running but tasks are not ready

Twins and task setup finish separately. Open Tasks, read any displayed error, and choose the action shown: When setup succeeds, Tasks shows Saved in Tasks with … scorers and Open task suite. If the draft has no executable tasks, choose Generate tasks, review them, and launch a version to save them. If it has unlaunched task or scorer edits, choose Launch version to include those edits.

​

My latest edits aren’t in the saved world

Edits stay in this browser as a draft until you launch. Choose Launch version to save them as a new version. Another browser sees the last saved launch. Clearing browser storage discards unlaunched drafts. Editing the draft does not change the running twins. Calling a twin’s API changes its runtime records, not the saved blueprint. Check which world version you are viewing before comparing its configuration and activity.

​

Timeline is empty

Check which activity you expect to see:

• The world’s Timeline shows requests and recorded lifecycle changes from its dashboard twins. Confirm the world is launched and the agent is calling the intended twin.
• An evaluation trial has its own fresh twins. Open that trial’s Timeline under Backtests to inspect its work.
• Use Load older activity to page back, and clear filters that hide the event. Filters apply to loaded activity.

If a request should have changed a record, select it and inspect its captured changes. A lifecycle event showing a twin start does not demonstrate that the agent completed a task. See Timeline and traces.

​

The agent cannot call a twin

In Twins, confirm the service is Running and use its current base URL and API token. A twin’s vendor API uses its own token. Evaluation trials receive their own service bindings; check the agent execution setup. For an authentication error, follow Authentication and read the specific error before changing credentials.

​

The agent runs but scores zero

Open the trial in Backtests and check these in order:

1. Task and agent version: confirm this is the trial you intended to run.
2. Actual request: did the agent call the service, or only describe an action in its answer?
3. Target and response: did the request reach the correct issue, case, order, or channel, and did it succeed?
4. Captured change: does the trial’s Timeline show the required update or new message?
5. Scorer: compare its target, expected value or text, and pass threshold with the captured result.

For the built-in task, the Slack check requires a new message containing ENG-2 is in progress for order 7001; case 00001001 remains open. A similar paraphrase does not satisfy that exact text check. The other scorer checks the update of ENG-2 to In Progress. World-state scorers need a trial’s captured evidence. The scorer editor’s sample test does not provide it. See Tasks and scorers and Read the results.