OpenAI Agents SDK Adds Sandbox Execution for Governance
OpenAI now offers sandbox execution in its Agents SDK. This feature lets enterprise governance teams run automated workflows while keeping risks under control.
Teams moving AI systems from prototypes to production often face tough choices about operations locations. Model-agnostic frameworks gave some flexibility at first. Yet they did not make full use of advanced models. SDKs from model providers stayed close to the models. Still, they often missed strong controls and visibility.
Managed agent APIs made deployment easier. But they limited where systems could run and how they reached company data. OpenAI addresses these issues with new Agents SDK features. Developers get standard infrastructure with a model-native harness and built-in sandbox execution.
Aligning Execution with Model Patterns
The new setup matches how models naturally work. This boosts reliability for tasks that span different systems. Oscar Health shows how it handles unstructured data well.
The healthcare company tested the infrastructure on a clinical records workflow. Past methods failed to manage it consistently. Engineers needed the system to pull accurate metadata. It also had to grasp patient encounter limits in detailed medical files. Automation speeds up parsing patient histories. That helps coordinate care and enhances member experiences.
Rachael Burns, Staff Engineer and AI Tech Lead at Oscar Health, said: "The updated Agents SDK made it production-viable for us to automate a critical clinical records workflow that previous approaches couldn't handle reliably enough. For us, the difference was not just extracting the right metadata, but correctly understanding the boundaries of each encounter in long, complex records. As a result, we can more quickly understand what's happening for each patient in a given visit, helping members with their care needs and improving their experience with us."
Deploying such systems means handling vector database updates, curbing hallucinations, and cutting compute costs. Without standards, teams build fragile custom links.
Tools and Orchestration Improvements
The model-native harness cuts these problems. It adds adjustable memory, orchestration that knows about sandboxes, and filesystem tools like Codex. Developers use standard parts: tool calls through MCP, custom rules in AGENTS.md, and file changes with the apply patch tool.
Skills enable step-by-step disclosure. The shell tool runs code. Standardization frees engineers from infrastructure tweaks. They can build business-specific logic instead.
Linking autonomous programs to old tech stacks needs exact routing. For unstructured data, retrieval pulls key context.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
The SDK adds a Manifest to standardize workspaces. Developers describe setups, mount local files, set output folders. Connections go to AWS S3, Azure Blob Storage, Google Cloud Storage, or Cloudflare R2.
A clear workspace tells the model input spots, output paths, and organization for long runs. It stops broad data lake searches. Access stays to checked contexts. Governance tracks decision sources from prototypes to production.
Sandbox Support and Security Measures
Sandbox execution comes built-in. Programs run in controlled setups with files and dependencies. No manual assembly needed. Use custom sandboxes or ones for Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop, or Vercel.
Enterprises worry most about risks in autonomous code. Security assumes attacks like prompt injection or data theft from external reads or generated code.
OpenAI splits the control harness from compute. Credentials stay out of code-run environments. Malicious code cannot reach the control plane or main keys. This blocks network-wide attacks.
The split also fixes failure costs. Tasks fail from timeouts, crashes, or limits. A 20-step financial report failing at step 19 wastes resources on restarts.
With this setup, sandbox loss does not end the run. State stays outside. Snapshotting and rehydration restart from checkpoints in new containers. This saves on cloud costs for long tasks.
Scaling uses dynamic allocation. Runs call one or many sandboxes by load. Subagents go to isolated spots. Tasks parallelize across containers for speed.
Availability and Future Updates
These features work for all API customers at standard token and tool prices. No special contracts needed. Python developers get them first. TypeScript comes later.
OpenAI will add code mode and subagents to both libraries. Plans include more sandbox providers and ways to link into internal systems.
OpenAI, known for models like GPT-4, builds tools for safe AI deployment. Oscar Health, a tech-focused health insurer founded in 2012, leads in using AI for better care.