Sparta: Stateful agent regression testing

Agent tests need state, not transcripts.

What it does

Sparta’s Worlds is an AI-agent test environment that creates a sealed copy of systems such as Zendesk, Stripe, Shopify or Salesforce. The homepage says it keeps state, runs the agent against the twin, then grades what changed against what should have changed. It names failures injected, database state and byte-for-byte replay as the evidence surface.

The pricing page lists a free plan with one world and unlimited runs on the customer’s machine. The Team plan is $400/month and adds every world in the catalog, several worlds in one environment, multi-agent chains and CI integration. These are dated page observations; they are not proof of test quality.

The docs make the integration boundary unusually clear. They describe a ten-minute path from local world to CI, but also say the current dashboard is not part of the published sparta_worlds@0.8.0 package used by the commands on the page. That caveat belongs in the buying decision.

Buyer and task

The buyer is an agent team that needs to know whether an action changed the system correctly, not merely whether the transcript sounds reasonable. The first task should be one stateful workflow with a known before-and-after: create a support ticket, update a billing record or change a CRM object, then grade the resulting state.

A useful test also includes a failure. Give the agent a permission error, a stale record or a malformed input. Check whether the world resets cleanly, whether the failure is reproducible and whether the test can run in CI without hidden local state.

The first test

Pick one workflow, define the expected state transition, run it three times and compare the diff. Then run the same test after a deliberate schema or permission change. If the system only shows a transcript, it has not solved the verification problem. If it shows the changed state, the expected state and a replayable failure, it has something a team can put in a release gate.

Sources checked

official Speedrun profile · company homepage · pricing · docs

Sources checked — September 20, 2026.