What is an agent sandbox?
An agent sandbox is an isolated execution environment where an AI agent runs: its own filesystem, network policy, and compute, separated from your production systems. The agent can execute code and use tools inside the boundary, while every interaction with the outside world passes through explicit, audited grants.
Sandboxes exist because agents act. A model that only writes text can be filtered; an agent that runs code and calls systems needs containment, so a compromised or confused agent damages a disposable environment, not your infrastructure.
Credential isolation is the companion rule: secrets never enter the sandbox. On Allocate, tools that need credentials execute outside the sandbox and return results into it, so a prompt-injected agent has nothing to steal.
See Cloud AgentsRelated terms
Allocate is the cloud inference platform for companies that want to train and run their own models.