Skip to main content
This walkthrough tests whether your agent can start engineering triage of a duplicate invoice and notify finance. You will launch a world, run its two tasks, and inspect the invoice task’s two scorer results. Before you run, you need a Chronicle account and a registered, runnable agent with the task’s tools and model credentials. Registration alone does not package or execute the agent. Follow Connect your agent if execution is not configured yet. You can explore the records and task without an agent in the interactive product example.
1

Launch the example world

Open Worldsmith and choose Coordinate an order-to-cash rollout. Chronicle loads connected Slack, Linear, Salesforce, and SAP records, plus two ready-made tasks.
Draft Linear issue ENG-2 with its duplicate-invoice description and cross-service references

The draft's ENG-2 record links order 7001, case 00001001, and the finance handoff channel. Launching uses these records as the starting data.

Open the issue record at full size.Choose Launch world. In the world’s Twins tab, wait for every service to show Running. Launching also saves the tasks as a suite.Checkpoint: open the world’s records and find Linear issue ENG-2. Its incident description references order 7001, case 00001001, and #erp-alerts—the same IDs used in the task.
2

Read the task and its checks

In the world, open Tasks → Open task suite. Select Triage the duplicate invoice incident. The agent must:
  1. Read ENG-2 and case 00001001 to confirm the incident.
  2. Move ENG-2 to In Progress.
  3. Post a new message in #erp-alerts containing this exact sentence:
ENG-2 is in progress for order 7001; case 00001001 remains open.
The two attached scorers check the issue change and the handoff message. Both have a pass threshold of 1. The handoff check matches the exact sentence, so a paraphrase can fail.The instruction also says not to close the case or claim a fix. Add a separate scorer if you want either constraint checked independently.
3

Run an evaluation

Open Backtests → New backtest and fill in the rows:Choose Run backtest. Two tasks and one agent produce two trials. Each gets a fresh world; changes made while exploring the dashboard do not become the trial’s starting state.
4

Read the scores and evidence

When the run finishes, open Results → Cases and find the invoice task. Switch to Raw trials and select its attempt. Read Scores for each check’s value and threshold, then follow the trial’s Timeline.Checkpoint: you can identify the agent version, its task, each scorer result, and the request or record change behind that result. A completed run alone is not a passing test.

Compare a second version

After the first run, register and configure an improved version, such as billing-agent@1.1.0. Create a Compare evaluation, select both versions, and choose Set baseline on 1.0.0. Keep the same task suite and starting world. Two tasks and two versions produce four trials. Run an evaluation explains the comparison and result views in detail.

Next steps

Strengthen the checks

Add a case-open guard or a judge for the agent’s explanation.

Build your own world

Describe the system and scenario you want to test.

Run from the API

Launch and inspect evaluations with your SDK.