The OpenAI Agent SDK is crazy! The "Harness/Compute Separation" architecture completely solves the security and scalability pain points of the long-range Agent. The🔥 OpenAI Agents SDK has been updated, and the "textbook-level" design of the enterprise-level Agent architecture has been directly disclosed. The core is the Harness and Compute separation architecture in this figure, which perfectly solves the two major problems plaguing Agent developers: security isolation & long-term stable operation. 🧩 Core architecture disassembly: This diagram breaks the Agent system into two completely independent but cooperative modules, with extremely clear responsibilities at each layer: • Left: The trust layer (Harness + Secrets) This is your "trust base", running in a trusted environment that you can control (supports any platform such as Temporal, AWS, Azure, etc.), and is not exposed to external sandboxes at all: ✅ Manage all sensitive credentials, API keys, and completely eliminate the risk of sandbox code leaking ✅ keys Agent Loop scheduling, MCPS/Tools Orchestration, deciding when to invoke the model, when to trigger the tool ✅ to interact with external databases, APIs, public networks, all high-privilege operations are done here • Right: Compute layer (Sandbox + storage) This is "isolated execution sandbox", specially run code, shell commands, file operations generated by the model: ✅ Support 7 mainstream sandbox environments such as OpenAI official sandbox, E2B, Vercel, etc. ✅ Support local/mounted file systems as needed, seamlessly docking S3, Azure Blob, GCP, etc. There are only execution permissions in the cloud storage ✅ sandbox, and there are no highly sensitive credentials. Even if it is injected into the attack, it will not cause too 🔑 much impact to the architecture. What pain points have been solved?
- Secure isolation: Lock the risk in a cage In the traditional architecture, where the Agent code runs, the key must be where, once attacked by Prompt Injection, the key is directly leaked. Now, the Harness layer is only in a trusted environment, and the sandbox layer is only responsible for execution. Even if the sandbox is broken, no core credentials can be obtained, perfectly realizing the "principle of least privilege".
- Persistence and long-term execution: After the separation of the Harness layer and the Compute layer, the state and progress of the Agent can be saved to the file system/cloud storage independently. Even if the sandbox restarts and crashes, it can recover from the breakpoint and support long tasks for hours or even days. This is not possible with the traditional Agent architecture.
- Portability Full: Anywhere can run No matter whether it is the Harness layer or the Sandbox layer, cross-platform deployment is supported. You can run Harness on-premises and sandbox in the E2B cloud environment; you can also deploy it all in AWS/Azure without being bound by a single platform.
- Debugging and auditing are simpler. All high-privilege operations and external requests are concentrated in the Harness layer. Logs and audit trails can be done uniformly here. The sandbox layer is only responsible for execution. When troubleshooting problems, it is not necessary to flip through the logs in a complex sandbox environment. What 💡 does this mean for developers? Simply put, OpenAI has made the pit they stepped on on the Codex and verified best practices directly into an out-of-the-box architecture: • No longer need to isolate themselves from the random sandbox, key management, and persist the state, the SDK has done it directly for you • You can rest assured that the Agent can run complex code and shell commands without worrying about inner leaks and environmental pollution • Long tasks finally have a reliable breakpoint running ability, and no longer have to start again because of one network fluctuation # OpenAI # AgentSDK # AI architecture # developer With # LLM