Repertoire

Use cases for agents that have to act.

Make your agents more capable in the real world — without putting the real house on the line.

Twins of Slack, Gmail, Calendar, and Stripe so a support agent can post, search, and refund against a seeded unpaid invoice.

Grade every commit or PR after the agent runs in Evalab. Judging criteria stay attached to the trace.

Short-lived or permanent twins so tool loops spend tokens against stateful understudies, not production APIs.

The same twin can speak HTTP, MCP, and CLI so the agent’s preferred door still opens onto the sandbox.