Designing execution boundaries for Agent tools
An Agent may need to read files, write results, or run tools. A useful sandbox design starts by deciding which work the Agent is authorized to perform and then limits what each tool can access while doing it.
Start with the work scope
Each AgentRun is bound to a tenant, workspace, and, when relevant, a work item. The platform resolves the working directory and tool policy from that authorized scope before execution. A tool request cannot choose a different tenant or widen its own workspace access.
Use more than one boundary
Application checks validate paths and tool permissions. Capabilities that execute commands can also run behind a system-level isolation boundary with a task-specific filesystem view and a limited environment. This separation keeps runtime configuration and credentials outside the tool's normal working view.
Treat network access as a separate capability
The Agent runtime needs network access for model and platform services. A tool's network access is a different decision and is governed by the applicable execution policy. This prevents a general-purpose tool from inheriting every connection available to the runtime controller.
Be precise about the boundary
Tool-level isolation does not mean the entire Agent process runs in one sandbox. Different tools have different execution paths and controls. Permission checks, tool isolation, audit records, and skill contracts work together; none of them alone defines the whole security model.