OpenAI turned its coding agent into general-purpose infrastructure — what the new Agents API means for AI product builds
OpenAI launched its Agents API in public beta on 10 September 2026, exposing the session orchestration, context compaction and sandboxed execution it built for its Codex coding agent as general-purpose infrastructure any developer can build on, billed through normal model and tool usage rather than a separate fee.
13 September 2026
On 10 September 2026, OpenAI put its Codex agent harness behind a public API. The Agents API hands developers the same machinery OpenAI built to run its own coding agent — durable sessions that continue across turns, automatic context compaction for long-running work, coordination across subagents, support for MCP and custom tools, and execution inside an OpenAI-hosted sandbox or a developer’s own compute environment. There’s no separate charge for the API itself; usage is billed through whichever underlying models and tools a session calls. OpenAI also shipped GPT-Live-1 to general availability the same week, adding full-duplex voice conversation that can run alongside a backend agent handling reasoning and tools.
Why this is bigger than a coding-tools story
The interesting part isn’t that OpenAI improved its coding agent — it’s that the hard infrastructure problems behind any AI agent (keeping a session alive over hours or days, compacting context so it doesn’t blow the window, running tools safely in a sandbox, coordinating several subagents on one task) are now something a team can call an API for, rather than something they have to build themselves. That’s the same trajectory cloud computing took: undifferentiated heavy lifting turns into a managed service, and the competitive question shifts from “can you build the plumbing” to “what does your agent actually do, and how well is it scoped to the problem.”
So what
If you’re a founder or product lead weighing whether to commission an AI agent product — not a chatbot bolted onto a website, but something that does multi-step work on a user’s behalf — this lowers the cost and risk of the infrastructure layer, but it raises the bar on everything else: the scoping, the guardrails, and the judgement about what the agent should and shouldn’t be trusted to do unsupervised. That’s the part that still needs a team who’s built agentic products before, not just a team that knows the API exists. Get in touch through our contact page if you’re scoping something in this space and want a straight read on what’s realistic.