OpenAI Developers
@OpenAIDevs
With the Agents SDK and @Vercel Sandbox, agents can execute work in isolated environments while keeping credentials separate from the harness. https://t.co/luR5oF05du
17retweets250likes
View on XOpenAI Secures Agent Workflows by Separating Credentials from Execution Environments
· Updated
OpenAI updated its Agents SDK to include native support for Vercel Sandbox, providing isolated environments for autonomous task execution. This integration allows agents to run code within secure MicroVMs (lightweight virtual machines that isolate processes) decoupled from core orchestration logic. Developers can now trigger these sandboxes on demand to handle specific execution steps.
This update addresses the primary security bottleneck in agentic engineering: the risk of executing agent-generated code. By separating the "harness" from the execution environment, sensitive credentials never enter the sandbox. This physical isolation ensures that if an agent's workspace is compromised, the core system and its secrets remain protected.
You can now parallelize workflows by routing subagents to their own dedicated sandboxes simultaneously. The SDK supports invoking sandboxes only when necessary, optimizing compute costs while maintaining strict security boundaries. These features are available through the latest Agents SDK and require a Vercel account for infrastructure management.
