Skip to main content
Synthetic Tool state belongs to an instance within a selected project and environment. Your application’s own state remains in your runner.

​
Independent runs

A recipe’s Tool bindings describe the builds, datasets, seeds, and operations needed by a case. The runner allocates separate instances for independent scenarios, configurations, and repetitions, then closes the run-owned instances. A run does not share another run’s synthetic records merely because both use the same Tool build.

​
Continuing simulations

A scenario sequence keeps synthetic state across its steps. Your agent can act, observe consequences, advance modeled time, and act again within that controlled sequence. Repeated trials are separate from continuing steps: repetitions start independent runs; sequence steps continue the same run.

​
What isolation does not mean

A native callback can still share your process. A command can still access your filesystem. Your HTTP application can still have external state. Synthetic Tool isolation does not automatically reset the agent’s memory, your database, OS timers, or model randomness. Make application reset hooks explicit and use prepared isolated runners when that boundary is required. See state and time and security.