Skip to main content
A task is a reproducible unit of agent work: a goal, the environment it runs against, a persona, and the verifiers that grade it. Tasks are what you run in evals and what you publish for others in your org to reuse.

Task shape

matching_policy is the bridge between tasks and policies: running the task drives the latch, so you can watch the exact verdict it produces in a controlled run.

Publish

1

Add the publishing skill (one-time)

2

Publish the current task bundle

From the console (Tasks → Publish) with the slash command:
or from the CLI:
3

Result

Publishing versions the bundle (version + rev), uploads its verifiers, and makes the task forkable by other agents and teams in the workspace.

Legacy simulation tasks

Older simulation scenarios use a lighter shape (scn_001, …):

Next: provision an environment

Give tasks a sandboxed world to run against.