OpenAI Agents SDK update
OpenAI released updates to its Agents SDK that add sandboxed execution, evaluation tooling and a separation between the orchestration 'harness' and the compute that runs tasks. The change is described as making agents easier to test, observe and govern by isolating how workflows are orchestrated from where they execute. (techcrunch.com)
OpenAI updated its Agents software development kit on April 15, adding native sandboxes and a new “harness” layer for running agents more safely. (openai.com) In plain terms, agents are software systems that use a model to plan steps, call tools, and complete multi-step jobs instead of answering a single prompt. OpenAI’s developer docs say the kit already handled handoffs between specialized agents, tracing, and tool use before this release. (developers.openai.com) The new harness is the part that coordinates work, like a dispatcher assigning jobs, while compute is the machine or environment that actually does the work. OpenAI said separating those layers lets developers keep orchestration logic while changing where tasks execute. (openai.com) The other big addition is sandbox execution, which gives an agent a contained workspace with files, commands, packages, ports, snapshots, and memory. OpenAI’s docs say that feature is available now in the Python Agents software development kit. (developers.openai.com) OpenAI said the updated kit is meant for agents that work across files and tools on a computer, not just chat windows. The company described the release as infrastructure for “secure, long-running agents” that can keep working through larger tasks. (openai.com) Evaluation is part of the pitch because companies need ways to test whether an agent actually finishes jobs reliably before giving it real work. OpenAI’s cookbook has already been steering developers toward online and offline evaluation, plus tracing of each internal step, as teams move agents into production. (developers.openai.com) The timing fits a broader shift in OpenAI’s platform from single responses toward tool-using systems that can act inside software environments. Its “building agents” track says the kit is designed for workflows with tools, orchestration, tracing, and guardrails rather than one-off model calls. (developers.openai.com) OpenAI said the new harness and sandbox launch first in Python, with TypeScript support planned later. The company also said code mode and subagents are still in development. (helpnetsecurity.com) TechCrunch reported the update as part of a race among OpenAI and Anthropic to sell enterprises the tooling to build automated helpers on top of frontier models. This release moves that competition further into the engineering layer where companies care about testing, monitoring, and control. (techcrunch.com)
