OpenAI Agents SDK update brings production-ready AI tools
OpenAI Agents SDK update adds sandbox execution and a new harness to help developers build reliable, production-ready AI agent systems.
AI agents are growing up. And they need infrastructure.
OpenAI announced the next evolution of its Agents SDK, positioning it as standardised infrastructure to help developers move agent-based systems from prototype to production.
The update introduces a model-native harness and built-in sandbox execution, aiming to reduce the engineering overhead that often slows down real-world deployment. This is less about building smarter agents. It is about making them deployable.
From experiments to production systems
Most agentic systems today work well in demos but struggle in production. The gap lies in orchestration, memory, execution environments and reliability. OpenAI’s updated Agents SDK is designed to bridge that gap.
Instead of stitching together custom tools and workflows, developers now get a structured framework that handles how agents plan, act and execute tasks across files and tools. This reduces the need for what engineers often call “plumbing”, the custom glue code that connects different parts of a system.
A harness that mirrors how models actually work
At the core of this update is a redesigned harness. The harness is essentially the system that manages how an agent operates. OpenAI says the new version aligns more closely with how modern AI models naturally process tasks, which improves performance on complex and long-running workflows.
It introduces features like configurable memory, filesystem utilities and sandbox-aware orchestration. These allow agents to manage files, run commands and maintain context over extended tasks.
The SDK also standardises common agent patterns. This includes tool usage via the Model Context Protocol, structured file edits through an apply-patch system, command execution using a shell tool, and custom instructions through an AGENTS.md convention.
For developers, this means fewer decisions about how to structure an agent. The framework handles the basics, allowing teams to focus on what actually matters: the product experience.
Built-in sandboxing for safer execution
One of the biggest additions is native sandbox execution. Agents often need to run code, install dependencies or interact with files. Doing this safely has always been a challenge. The updated SDK solves this by providing controlled environments where tasks can run without risking the broader system.
Developers can either bring their own environments or use built-in support from providers like Cloudflare, Vercel and others. A key part of this system is the Manifest abstraction. This defines the agent’s workspace, including where inputs are stored and where outputs should go.
It gives the model a predictable structure, which improves reliability during long tasks. In short, the agent always knows where to look and where to write.
Security, durability and scale built in
Production systems need more than functionality. They need reliability and safety. OpenAI has separated the harness from compute, ensuring that sensitive credentials are not exposed to model-generated code. This helps reduce risks like prompt injection and data leaks.
The SDK also supports durable execution. This means agent states can be saved and restored if something fails, using techniques like snapshotting and rehydration. For scaling, teams can distribute work across multiple environments. Subagents can run in isolation, tasks can be parallelised, and compute resources can be used only when required.
This makes the system more efficient and easier to manage at scale.
Google rolls out Personal Intelligence in the Gemini app for users in India
The updated capabilities are available through OpenAI’s API under standard token and tool-based pricing. The new harness and sandbox features are currently live in Python, with TypeScript support expected soon.
OpenAI has also hinted at upcoming additions such as code mode and more advanced subagent capabilities, signalling that this ecosystem will continue to evolve.
The latest Agents SDK update signals a clear transition in the AI landscape. The focus is moving from experimentation to execution. By standardising how agents are built, run and scaled, OpenAI is addressing one of the biggest barriers to adoption.
If these tools deliver on reliability and ease of use, developers may spend less time wiring systems together and more time building meaningful applications. Because the future of AI agents will not be defined by what they can do in isolation, but by how well they work in the real world.