OpenAI launched its Agents API on September 10, a managed service for building and running cloud agents powered by the same Codex harness the company uses internally S¹P³. The API handles orchestration, long-running sessions, and tool use out of the box: the three things that have made agent development painful to build from scratch. What remains unclear is whether a managed service can beat the open-source frameworks already gaining traction, and at what cost.

My read: This is OpenAI doing what it does best: taking something it built for itself, hardening it at scale, and selling it as infrastructure. The Codex harness has been running for millions of users through Codex and ChatGPT for Work P³, so the engineering is battle-tested in a way most agent frameworks aren't. But the evidence pack contains zero pricing information and no general availability date, which means I can't assess whether this is competitive with what's already free. The open-source alternatives, Hugging Face's smolagents with 29,000 stars on GitHub P⁶ and OpenAI's own open-weight Agents SDK P⁴P⁵, aren't going away. Whether managed convenience beats self-hosted control is the trade-off that will decide this.

What the Codex harness actually does

The Agents API runs on the Codex harness, the same engine OpenAI built to power its own coding agent S¹P³. According to OpenAI's developer documentation, agents built with the SDK design their actions, utilize external utilities, work with other specialized agents, and retain sufficient context to finish complex, multi-stage tasks P⁵. That last phrase is the important one. Multi-step work, tasks that require an agent to remember what it did three steps ago, call an external tool, adjust its plan, and keep going, is exactly where most DIY agent setups fall over.

The API bundles three capabilities that developers previously had to wire together themselves S¹:

  • Orchestration: coordinating multiple agents or steps in a single workflow
  • Long-running sessions: maintaining state across extended tasks without timing out or losing context
  • Tool use: letting agents call external functions, APIs, and services

OpenAI's own description says the company scaled Codex and ChatGPT for Work to a massive global user base and learned what infrastructure that requires P³. The Agents API is that infrastructure, packaged.

How it fits alongside what already exists

OpenAI already offers an open-source Agents SDK, a Python framework for multi-agent workflows that lives on GitHub P⁴. The developer docs describe it as a way to create agents using standard programming and scale up to more sophisticated execution models when necessary P⁵. The Agents API appears to be the managed, cloud-hosted counterpart: you write agents the same way, but OpenAI runs them.

This matters because the alternative, running your own agent infrastructure, means handling session persistence, tool execution, error recovery, and orchestration logic yourself. Hugging Face's smolagents library, with 29,038 GitHub stars and an Apache 2.0 licence P⁶, takes a barebones approach: agents that reason using programming logic with minimal abstraction. It's popular because it's simple and free. But it doesn't run in the cloud, and it doesn't manage sessions for you.

The trade-off is the familiar one in cloud computing: control versus convenience. OpenAI is betting that enough developers want the convenience of a managed service, with no servers to provision and no session management to build, to pay for it.

What to do about it

If you're building agents today, the Agents API gives you a new option: write your agent logic, hand it to OpenAI, and let them run it. The practical appeal is strongest for teams that have tried building agent infrastructure and hit the wall on session management or orchestration.

Consider a logistics startup that wants an agent to monitor shipment data across three carrier APIs and flag delays, then email customers with revised ETAs. Today, that team would need to build the orchestration layer, handle API authentication for each carrier, manage session state so the agent remembers which shipments it has already checked, and deal with retries when a carrier API times out. The Agents API promises to handle orchestration and session management S¹, so the team can focus on business logic.

The open-source Agents SDK P⁴P⁵ remains the starting point. The developer docs recommend beginning with the SDK quickstart and growing into more advanced patterns. If your agents are simple, smolagents P⁶ or the SDK alone may be enough. The managed API earns its keep when sessions get long and orchestration gets complex.

One practical step this week: read the Agents SDK guide on OpenAI's developer site P⁵ and build the quickstart agent. That gives you a baseline to compare against when pricing for the managed API is announced.

What we don't know yet

The announcement is fresh, dated September 10, 2026, and several critical details are missing from the evidence:

  • Pricing: no cost information appears in any source. Whether the managed service is priced per session, per tool call, or per token is unknown.
  • General availability: no GA date is specified. The announcement says "introducing" but doesn't clarify whether this is a limited preview or broadly available.
  • Regional availability: no information on data residency, regional hosting, or availability outside the US.
  • Independent benchmarks: no third-party testing or comparison against existing frameworks is available yet.
  • Model support: the API is powered by the Codex harness S¹, but the sources don't specify which underlying language models are used or whether developers can choose.

The next signal: OpenAI's developer conference or next API pricing update, where managed service costs and availability tiers are typically announced.

Sources: S1 — Introducing the Agents API · S2 — Introducing the Agents API - OpenAI · P3 — Introducing the Agents API | OpenAI · P4 — GitHub - openai/openai-agents-python: A lightweight, powerful framewor · P5 — Agents SDK | OpenAI API · P6 — huggingface/smolagents

More from Not A Tech Guy

Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.