sandtable

sandtable rehearses your agent before it touches real customers. Key capabilities: • Simulated callers that test your agent's conversations • Stateful mock systems that hold real records and balances • Every business rule checked on every tool call • Deterministic grading into pass^k reliability scores • Open source CLI with no judge model An agent can sound fine on the call and still book a carrier whose insurance has expired. sandtable catches failures like this by checking the actions an agent takes, not just how it talks. Each trial runs against a mock world with real state, so a booking, refund, or update either follows the rules or fails the run. Because grading is deterministic, the same trial produces the same result every time, and pass^k measures whether an agent succeeds consistently across repeated runs. Domains are defined in a spec, and an init guide helps draft them. The library and HTTP protocol let you run against your own agent, including agents that call their own tools. Try it now with npx sandtable-ai demo. Built for teams shipping customer-facing agents in operations, logistics, support, and sales who need evidence before release. Use it locally during development or in CI to block regressions, and share results with customers for sign-off.