OpenAI just made your orchestration layer a cost center
On September 10, OpenAI put its Agents API into public beta. It is not a new model. It is the plumbing behind Codex and ChatGPT for Work, opened up so any developer can build on it: session continuity, context compaction, tool routing, subagents that run in parallel. And the pricing is blunt. There is no markup. You pay for tokens, tools, and sandbox time. The orchestration itself is free.
That single fact should make a lot of engineering leads uncomfortable, because a lot of teams spent 2025 and early 2026 building exactly this layer by hand.
What's actually in the box
Strip away the announcement copy and three things matter.
First, context management that compacts a session automatically as it nears the model's context limit. No more writing your own summarization loop every time a task runs long.
Second, subagents. You can split a task into independent pieces, hand them to parallel workers, and let the main agent merge the results. That is the pattern most teams have been hand-rolling with queues and custom state machines.
Third, a choice of execution environment. Run it in an OpenAI-hosted sandbox, on your own infrastructure, or through a partner like Cloudflare, Modal, E2B, or Vercel. That third option matters more than it looks: it means you are not locked into OpenAI's compute even if you're locked into their harness.
The build vs buy call
Here's the uncomfortable part. If your team spent the last two quarters building a custom orchestration layer, that work solved a real problem at the time. It's just not a differentiated problem anymore. Context compaction, session state, tool call routing: these are becoming table stakes that a platform vendor gives away to sell you tokens. Maintaining a bespoke version of that stack going forward is time spent on infrastructure your competitors get for the price of an API key.
We'd tell a client the same thing we're saying here: don't rip out a working system on launch day, but stop investing further engineering time in commodity plumbing. Redirect that time toward the things a managed API cannot give you.
What still needs a human hand
The orchestration layer was never the hard part of building a good agent product. The hard part is knowing when the agent is wrong. Evals that catch a bad tool call before it hits a customer. Guardrails that stop an agent from taking an action it shouldn't, even when the model is confident. Business logic that encodes what "done" actually means for your specific workflow, not a generic one.
None of that ships in a managed API, and it shouldn't. It's specific to your product, your data, and your users, and it's exactly where a vendor's incentives diverge from yours. OpenAI wants you shipping more agent calls. You want agents that fail safely and rarely. Those are related goals, not the same one.
There's also a real infrastructure decision buried in the sandbox choice. Data residency, network topology, and credential handling are not solved by picking "hosted" by default. Teams in regulated industries especially need to treat that as an architecture decision, not a checkbox.
The takeaway
Commodity infrastructure getting cheaper is good news if you act on it. It's bad news if you keep funding a team to rebuild what a platform now gives away. The move this quarter isn't to chase every new managed API. It's to audit what your team is spending engineering time on and ask honestly whether it's still worth owning.
We're here to help founders and teams design and build digital products that are built to scale with you, not slow you down. If you're looking to build something, get in contact with us today!