The next evolution of the Agents SDK
The updated Agents SDK helps developers build agents that can inspect files, run commands, edit code, and work on long-horizon tasks within controlled sandbox environments.
TL;DR
- The updated Agents SDK enables developers to build AI agents capable of inspecting files, running commands, and editing code in controlled sandbox environments.
- New features include a model-native harness with configurable memory and sandbox-aware orchestration, along with native sandbox execution support for running agent work safely.
- The SDK offers standardized integrations with primitives like tool use (MCP), progressive disclosure (skills), custom instructions, and code execution (shell, apply patch).
- Developers can utilize built-in support for various sandbox providers or bring their own, with a Manifest abstraction for describing the agent's workspace.
- Separating harness from compute enhances security, durability (via snapshotting and rehydration), and scalability of agent systems.
- The new capabilities are initially launching in Python, with TypeScript support planned for a future release.
- Customers like Oscar Health have used the updated SDK to automate critical clinical records workflows.