openai.com
|
ksl
|
|
OpenAI shipped a significant update to its Agents SDK, adding sandboxing and a new harness system for long-running, multi-step tasks. Sandboxing lets agents operate in isolated environments with scoped file and code access – a direct response to enterprise concerns about letting AI agents touch production systems. The harness piece is arguably more interesting: it gives developers infrastructure for agents that persist across complex workflows instead of running single-shot. Python support comes first, with TypeScript following later. Anthropic, LangChain, and CrewAI have all been building in this same direction, but OpenAI packaging sandbox isolation at the SDK level narrows the gap between prototype and production-grade agent deployments.
