Analysis
OpenAI introduced the Agents API in public beta on September 10, exposing the same harness and infrastructure that powers Codex -- its own coding agent -- to any developer through a single API, the company announced. The pitch is narrow but useful: building a reliable agent requires a harness that manages context, calls tools efficiently, and coordinates subagents, plus infrastructure that keeps sessions running for days across files, code execution, and saved intermediate state. Most teams rebuild a version of that plumbing from scratch; OpenAI is now renting out its own.
The API is built from four objects -- an Agent (model, instructions, tools, and MCP servers), an optional Environment sandbox, a durable Session, and the stream of events a session emits back to the caller. Inside a session, the API can run code, edit files, search the web, apply predefined skills, produce artifacts, and split work across subagents under a set concurrency limit, MarkTechPost detailed.
“A framework startup whose only differentiation is session management and sandboxing now has a much harder pitch.”
Developers choose where the agent's compute actually runs: an OpenAI-managed sandbox, their own infrastructure, or one of nine partner sandboxes -- Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel. That partner list is notable on its own: it puts OpenAI in the position of routing workloads to infrastructure providers, including Oracle, which is simultaneously one of OpenAI's own largest compute suppliers.
Pricing adds no separate API fee beyond model tokens and paid tools -- containers run at standard compute rates -- which lowers the barrier for teams that were priced out of building agent infrastructure themselves but still pay OpenAI's model rates either way. The company's own prior coverage on Pulse has tracked a steady cadence of infrastructure moves like this one, each aimed at making it harder for a developer to build a serious agent without touching an OpenAI-controlled layer somewhere in the stack.
The competitive read: this is OpenAI competing directly with the growing field of agent-framework startups -- LangChain, CrewAI, and a wave of smaller orchestration layers -- by commoditizing the exact harness those companies charge for. A framework startup whose only differentiation is session management and sandboxing now has a much harder pitch.
