OpenAI opened its Agents API to public beta on September 10, 2026 — and for any organisation that has been waiting for a reliable, managed way to run AI agents in production, this is worth paying attention to.

The API wraps the same harness and infrastructure that powers Codex, OpenAI’s internal coding agent, and makes it accessible to any developer through a single API call. You define the task, model, tools, and environment. OpenAI handles everything else.

What the Agents API Actually Does

The central problem with production AI agents has never been the model. It’s been the plumbing: keeping agents running across long tasks, managing context as conversations grow, recovering when something goes wrong, and coordinating multiple sub-agents without everything collapsing.

That’s exactly what the Agents API is designed to solve.

Under the hood, it provides:

Durable sessions — Agents can continue work across multiple turns, picking up where they left off rather than starting fresh each time. For multi-day tasks (a common pattern in real business workflows), this is a prerequisite.

Context compaction and recovery — OpenAI manages context windows automatically. When an agent runs long enough that it would otherwise hit token limits, the harness compacts the context intelligently rather than failing.

Orchestration for sub-agents — Agents can delegate to other agents. If a top-level task needs to fan out into parallel workstreams, the harness handles the coordination so developers don’t have to build that scaffolding themselves.

Built-in sandbox execution — Agents can run code, edit files, and save intermediate results inside OpenAI-hosted sandboxes. Developers can also connect their own infrastructure or a supported partner sandbox instead.

MCP server connections — The API connects natively to Model Context Protocol servers, which means agents can reach external tools and data sources in a standardised way.

The whole thing is available in public beta to any developer. Pricing follows the same model as the rest of the OpenAI API: you pay for tokens and tools used. No additional platform fee on top.

Why This Matters for Business

Most enterprise AI discussions treat “agents” as a future capability. The honest answer is that agents are already in production at many organisations — they’re just harder to build and maintain than anyone initially expected.

The infrastructure challenge is real. Keeping an agent session alive for hours or days, handling failures gracefully, giving agents access to the right tools without giving them access to everything: these are non-trivial engineering problems. Most teams have been solving them by hand, which is slow and fragile.

What OpenAI has done here is commoditise that infrastructure layer. Rather than each team building their own session management and recovery logic, they can now inherit it from the API. That should reduce the time to a working, production-grade agent from weeks to days.

This also changes the competitive landscape for AI agent platforms. Services that have been selling orchestration infrastructure now have to compete with the model provider itself offering the same thing at cost.

What This Means for Business

If you’re evaluating whether to invest in AI agents for your business operations, the Agents API shifts the calculation.

The barrier to production deployment just dropped. A team with API access can now build an agent that runs reliably for days, handles tool calls, manages its own context, and recovers from errors — without custom infrastructure. The scope of what’s practical has expanded significantly.

The build vs. buy question gets sharper. Specialist agent platforms now need a clear answer to “why not just use the Agents API?” For companies evaluating vendors in this space, that’s a useful pressure test to apply.

Governance still matters. The API gives developers more control over sandbox environments, which is positive. But as with any agentic system, the policies about what agents are allowed to do — and how you detect when they go outside their intended scope — remain your responsibility. A recent Enterprise Management Associates report found that 65% of enterprises had seen AI agents act outside their intended scope. Infrastructure doesn’t fix that. Policy does.

Cost visibility is cleaner. Pay-for-usage on tokens and tools makes it easier to model the cost of an agent-based workflow before you build it. That’s a practical advantage for anyone doing business case work.

For organisations that have been holding off on agents because the infrastructure felt too uncertain, September 10 is a reasonable moment to revisit that assessment.

---

Enterprise DNA helps businesses build AI agent systems and data capabilities that actually work in production. If you want to understand how agentic AI fits into your operations, book a discovery call with the team.