OpenAI abandons model-first focus, pivoting to orchestration of tools and workflows

OpenAI’s latest push is not a bigger model launch. It is a set of products for running agents across tools, files, and long-lived work. (openai.com) On April 15, OpenAI said its Agents SDK was getting a “model-native harness” and native sandbox execution. The company described the harness as the layer that manages instructions, tools, approvals, tracing, handoffs, and resume logic, while the sandbox handles isolated compute. (openai.com 1) (openai.com 2) That is a shift from treating the model call as the product. OpenAI’s own developer docs now define agents as applications that plan, call tools, collaborate across specialists, and keep enough state to finish multi-step work. (openai.com) The plumbing matters because a model alone cannot safely edit files, run commands, search the web, or wait for approvals. OpenAI’s Responses API now bundles built-in tools including web search, file search, computer use, code interpreter, and remote Model Context Protocol servers. (openai.com 1) (openai.com 2) OpenAI extended that idea into ChatGPT on April 22 with workspace agents, which it called “an evolution of GPTs.” The company said these Codex-powered agents can run in the cloud, keep working when a user is away, and be shared across a company in ChatGPT or Slack. (openai.com) The product language has changed with the releases. OpenAI tells developers to use raw client libraries for direct model requests, but to use the Agents SDK when the application owns orchestration, tool execution, approvals, and state. (openai.com) The same pattern shows up in Codex. OpenAI’s Codex cloud docs say coding tasks can run in the background, including in parallel, inside cloud environments tied to repositories and pull requests. (openai.com) OpenAI is also standardizing the contracts around outside systems instead of forcing everything through prompts. Its tools docs say models can load deferred tool definitions at runtime and connect to third-party services through connectors and remote Model Context Protocol servers. (openai.com 1) (openai.com 2) That makes the engineering problem look less like “pick the smartest model” and more like “design the workflow around failure.” OpenAI’s sandbox guide says orchestration can stay in a company’s own infrastructure while stateful execution runs in a separate sandbox for isolation and durability. (openai.com) OpenAI has not stopped shipping models, but its April releases put the emphasis elsewhere: who controls the tools, where the work runs, and how an agent resumes after something breaks. (openai.com)