Sandboxes
person concept tool org talk claim — click a node to jump to its page; hover an arrow for the relation
A sandbox is an isolated running environment used in agentic AI workflows to execute code, tools, or agent actions safely in response to events or on a periodic schedule.
Role in Agent Pipelines
Jason Lopatecki describes sandboxes as the execution context where an agent operates: "really the sandbox is this running environment where based upon an event or a periodic [trigger]" — 12:21. In the self-improving agent architecture he presents, the sandbox provides the controlled space in which the agent can take actions (such as running evals, modifying code, or opening pull requests) without risk to production systems.
Alternatives and Deployment Options
Lopatecki positions sandboxes as an alternative to full VPC deployment, noting that practitioners can choose from multiple providers depending on comfort and requirements: "you can use Arise sandboxes or you can use Daytona or any of these that you're comfortable with" — 22:09. This framing treats sandboxing as a modular, swappable infrastructure concern rather than a fixed architectural requirement, giving teams flexibility in how they isolate agent execution.