Skip to main content
An environment is a saved definition of the tools and starting records an agent gets during a trial. A published version of an environment is frozen, so every trial, today or next month, starts from the same state. You usually don’t create environments yourself. Launching a Worldsmith world publishes one automatically and binds it to the world’s tasks.

​
Draft, environment, twins

  • Worldsmith draft: the simulated system and scenario you are editing.
  • Environment version: a frozen recipe for the tools and records. Publish a new version when the starting state needs to change.
  • Running twins: live copies created from a version. API writes change their records during exploration or a trial.
Editing the draft never changes running twins, and calling twins never changes the saved version.

​
Bind a task to a version

A task’s environment answers “which starting state does this test need?”
  1. Open the task in Tasks.
  2. Expand Environment and select the environment.
  3. Select a specific published Version and wait for the saved state. Picking a different environment clears the version, so check it after changing.
For the duplicate-invoice task, pick the version where ENG-2 is still in Todo. A version where the incident is already resolved would test something else.
Task Environment controls with the example environment selected and Version set to worldsmith-v1.

The duplicate-invoice task is bound to a specific published version, worldsmith-v1, of its environment.

Open the environment controls at full size.
Always pick the version, not just the environment. The name alone doesn’t pin the starting state.
Tasks in one suite can use different versions. A task’s own version wins over an evaluation’s default environment; the default only applies to tasks that don’t have one.

​
Change the scenario without breaking history

Suppose the first version is too easy because the finance channel spells out the answer.
  1. Edit the Worldsmith draft and choose Launch version. This publishes a new environment version.
  2. Point the task at the new version (or let the new world’s suite do it).
  3. Run the agents again.
The old version and every result that used it stay as they were.

​
Custom environments

The Environments page lists saved environments and their versions. A custom version can describe recorded HTTP replays, seeded datastores, MCP servers, or standalone twins. To use an existing custom environment, select its published version in the task’s Environment section. Checkpoint: the task shows the intended environment and a specific version, and its saved state is confirmed. Run an evaluation to test the agent from that starting state.