Skip to main content
Worldsmith builds a world: a simulated system with services, data, and state your agent can interact with. This walkthrough models a company’s duplicate-invoice incident: move ENG-2 to In Progress and notify finance in #erp-alerts. Start with the example → Review the checked draft → Launch → Evaluate. Checking validates the draft. Launching saves it and starts its service twins. The agent does its work when you run an evaluation.
1

Open the example

In the dashboard, open Worldsmith and choose Coordinate an order-to-cash rollout. This loads the built-in company, connected records, and tasks used below.
Worldsmith welcome with its prompt composer and Coordinate an order-to-cash rollout starting point

Choose Coordinate an order-to-cash rollout to load the built-in example.

Open the welcome image. To create a different situation after this walkthrough, use a complete brief.
2

Follow the connected records

In Overview, review the company and shared people. In Twins, open Linear → Issues → ENG-2 and read the duplicate-invoice description.Check three references in the same incident:
  • SAP order 7001 is the order that was invoiced twice.
  • Salesforce case 00001001 tracks Northwind’s open customer case.
  • #erp-alerts contains the engineering and finance discussion.
Expanded Linear issue ENG-2 showing the duplicate-invoice description and its order, case, and Slack references

ENG-2's description connects the duplicate invoice to order 7001, case 00001001, and #erp-alerts.

Open the record image, or try the world workspace. In that example, use Overview to inspect people and Twins to expand Linear → Issues → ENG-2. Edits there stay in the example.Wait for World prepared in the Worldsmith conversation. Expand its notes to review any warnings. If you see A few things need fixing, correct the listed records or choose Fix with Worldsmith.
3

Read the task and its two checks

Open Tasks → Triage the duplicate invoice incident. Read the Instruction, Steps, and Expected outcome. The agent must move ENG-2 to In Progress and post a new message in #erp-alerts containing this exact text:
Triage the duplicate invoice incident with its instructions, steps, expected outcome, and two attached Python scorers

The built-in triage task has two Python scorers: one for the Linear update and one for the new Slack handoff.

Open the task image, or inspect the example task.Open each scorer. ENG-2 was moved to In Progress checks the issue update. The incident handoff was posted in #erp-alerts checks a new message containing the exact sentence above. Both have a 1.00 pass threshold.The instruction also says to keep the case open and not claim the bug is fixed. The two supplied scorers do not independently verify those constraints. To make them affect the score, add checks before launching; see grade constraints as well as changes.
4

Launch and confirm the task suite is ready

Choose Launch world. The button becomes available when checking has finished, there are no problems, and the draft includes at least one service.Launching saves this world version, starts its twins, and prepares its task suite and scorers. Confirm both checkpoints:
  • In Twins, the four services show Running. Each running twin has a base URL and an API token for its vendor API.
  • In Tasks, you see Saved in Tasks with … scorers and Open task suite. Task setup finishes separately from twin startup.
If Tasks shows Create task suite, Resume task setup, or Retry task setup, follow the task-setup recovery steps.The world’s Timeline records requests, changes, and lifecycle events from these twins. Once a request changes a record, select it to inspect the captured change. See Timeline and traces.
The local Linear twin's recorded mutation, showing state_id changing from an ID ending 0002 to one ending 0003 and the associated timestamp changes.

A real direct API update to the local Linear twin, not an agent evaluation. The stored state_id moved from Todo (ID ending 0002) to In Progress (ID ending 0003), as confirmed by the API response.

Open the recorded change at full size.
5

Run the task against your agent

Choose Tasks → Open task suite to review the saved tasks. If this is your first agent, register a runnable version, such as billing-agent@1.0.0. Then run the suite in Backtests.Every trial starts from a fresh copy of the saved world. Inspect the trial’s scores and Timeline to confirm the Linear update and Slack handoff. Its activity belongs to the trial, separate from the dashboard world’s Timeline.

Refine the next version

Reopen the world with the switcher beside its title. Click an editable field to change it; edits save to your browser’s draft and trigger another check. For a larger variation, ask Worldsmith:
Use Add twin, Remove twin, or Restore twin in Twins to change the services in the next version. Ask Worldsmith to change IDs, handles, or channel names so related references stay consistent. In Tasks, use Generate tasks when there are none, or Generate scorers / Add scorer for the selected task. Choose Launch version to save edits as a new version, or Relaunch for a fresh version of the same configuration. The v1, v2 controls open version history. Earlier versions keep their records and activity.
Draft edits stay in this browser until you launch. They do not change running twins. Screenshots use the built-in example; worlds generated from your own briefs can have different records and identifiers.

Write a good brief

Define the starting evidence, actions, and checks.

Worlds API

Compile and launch worlds from code.