OpenAI's Agents API turns the Codex harness into a managed service, and hand-rolled agent scaffolding gets harder to justify
OpenAI's Agents API entered public beta on September 10, putting the managed Codex harness (session state, tool orchestration, subagents, auto-compaction) behind one API call. Compute runs on OpenAI, your own infrastructure, or partner sandboxes like Vercel and Cloudflare.
OpenAI released its Agents API in public beta on September 10, and the short version is that the scaffolding most teams have been writing by hand to run an agent is now something OpenAI runs for you. A single call takes a task, a model, a set of tools, and a place to execute, and OpenAI manages the loop in between: planning, calling tools, spawning subagents, and keeping the session coherent across long runs. It is the same harness that powers Codex, exposed directly rather than bundled inside a product.
What the API actually does
The API is built around four pieces: an agent (model, instructions, tools, and MCP servers), an environment (an optional sandbox for file access and command execution), a session (a durable running instance that manages tasks), and the events that flow through it. The parts a team usually hand-builds are folded into the managed harness. Sessions auto-compact so a long run does not overflow the context window and force you to write your own summarization logic. Tool definitions load through a search step rather than getting stuffed into the prompt all at once, which trims token cost on tool-heavy agents. Tool calls can run in parallel and be filtered programmatically, and a main agent can delegate to subagents that each keep their own context. None of this is new as a concept. What is new is that it is OpenAI’s code running the loop, and that the harness is an open-source version of the Codex code you can actually read.
Where the agent runs is now your call
The more consequential decision is that OpenAI separated the harness from the
place the agent executes. You get three options. You can run on OpenAI’s hosted
infrastructure, the same Codex and ChatGPT backend, with configurable files and
skills. You can self-host by running codex exec-server over a WebSocket with
scoped keys, keeping execution on your own machines while OpenAI orchestrates. Or
you can plug into a partner sandbox, with first-class integrations at launch for
Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and
Vercel. That list matters because it means the agent runtime is not locked to
OpenAI’s cloud. A team already on Vercel or Cloudflare can keep execution where
its data and services live, which is usually where the compliance and latency
arguments get won.
What it costs, and what beta still means
There is no separate fee for the API itself. You pay for tokens, tools, and container time at standard rates, which keeps the pricing story simple and removes one reason to keep maintaining a bespoke harness purely to control cost. The caveats are the ones a public beta carries. Data residency is US-only for now, and Zero Data Retention is not yet supported, so regulated workloads and teams with strict data-handling commitments should treat this as a preview to evaluate rather than a production dependency. OpenAI says it plans to refine the API on developer feedback before general availability, the usual signal that the surface may still move.
What it means for developers
The practical shift is that the agent loop is becoming infrastructure you rent rather than code you own. For most teams building an internal agent, the harness has been the unglamorous, expensive part: retry logic, context management, tool routing, subagent coordination, and the sandboxing to run untrusted tool calls safely. Moving that to a managed service lowers the cost of starting and, more importantly, the cost of maintaining. Early adopters cited concrete gains, with Ciridae reporting a roughly 4x latency reduction from built-in subagent support and SafetyKit reporting a 60 percent cut in case-review cost, though those are vendor-supplied figures from the launch and should be read that way. The tradeoff is dependence. You are handing the reasoning loop to OpenAI’s harness, which means less to build but also less to inspect and control when a run goes wrong. The open-source harness softens that, because you can read what it does, but running it as a managed service is not the same as owning it.
What to watch
The questions worth tracking over the next quarter are whether the managed harness holds up on messy real-world agents rather than clean demos, whether the partner-sandbox model actually keeps teams off OpenAI’s own compute or quietly funnels them onto it, and how fast Zero Data Retention and non-US residency arrive, since those gate the enterprise adoption OpenAI is clearly aiming at. The larger pattern is that the major model vendors are competing to own the agent runtime, not just the model. Google shipped Gemini Managed Agents earlier this year and Anthropic has pushed the same way. OpenAI putting the Codex harness behind one call is the most direct version of that bet yet: the model is table stakes, and the orchestration around it is where the next lock-in is being built. Stackmaven’s follow-up will revisit around December 12.