On Thursday I published a build log about an orchestrator I wrote for Amazon listing work — a skill that calls four other skills, produces eleven documents in a dated folder, and knows when to stop. It took an evening on top of the skills it calls. The post was mostly about the constraint sentences, because they always are.
The same day, OpenAI put the Agents API into public beta and made the orchestration half of that a managed service with no additional fee beyond usage (openai.com, September 10 2026). Sessions, context compaction, subagent coordination, tool loading, recovery — the parts I spent an evening on and the parts most "AI-enabled" agencies have been quietly billing against — are now four API objects and a config block.
So here's the operator question, which is the only one that matters: does your agency's AI line item still have a cost basis? For most of them the honest answer is that it never had one, and this week removed the last place to hide that.
What happened
OpenAI released the Agents API in public beta on September 10, 2026, exposing the managed Codex harness behind one API call. Four objects: an Agent (model, instructions, tools, MCP servers), an optional Environment sandbox, a durable Session that persists across turns, and the events it streams back. OpenAI handles context compaction as sessions approach token limits, tool search and loading, and subagent coordination. You supply tools and pick where it runs — an OpenAI-hosted sandbox, your own via codex exec-server, one of nine partner sandboxes, or none at all.
Pricing, stated plainly: there is no extra fee. You pay for tokens, tools, and container time.
Why most brand owners will read this wrong
The dumb take arrives in two mirror-image flavours and both will be in your inbox by Tuesday.
"One person can now do what an agency does." No. The Agents API removes the plumbing between a model and a task. It does not remove the part that costs money, which is somebody who knows that a bullet claiming "waterproof" on a water-resistant product is a return, a one-star, and a contribution to a rate that eventually hangs a badge on your listing. I have now written some version of that sentence in five build logs and I've stopped treating it as a coincidence. The workflow is the cheap part. The constraint sentences are the product.
"It's a developer release, ignore it." Also no, and this is the more expensive mistake. A frontier lab just published the reference architecture for how an agent runs against a live commercial workload, with the execution environment as a selectable field. Specs don't stay in engineering. By October that field is a procurement question, and the agencies that have thought about it will answer in a sentence.
What actually changes for a brand doing $200K a month
The "we built a custom agent pipeline" pitch just lost its moat. Not its value — its moat. If an agency's differentiation was that they'd wired session management, retries and subagent fan-out, that is now a managed service available to every one of their competitors at no additional fee. Anything built on top of it that still differentiates is either proprietary data or somebody's judgment written down as rules. Ask which.
Your AI cost line does not move, and I'm holding that position. I've said since July that inside a $6,000 retainer the attributable model spend is low hundreds and sometimes double digits. Nothing this week changed it. Orchestration was never the expensive part — a managed harness that costs nothing extra saves you nothing, because you were paying nothing for it. What you pay for is the hour a senior person spends reading the output before it touches your catalogue, and that hour did not get cheaper on Thursday. If a vendor's rate card moves either direction on the back of this release, ask them to show you the line it moved.
The throughput ceiling is real and it's low. The published example caps concurrent subagents at three. Any pitch that implies a swarm of forty agents working your catalogue in parallel is describing something other than this product. Three at a time against 140 ASINs is a queue, and a queue has a duration, and a duration is a project plan somebody should be able to show you.
Two beta limits belong in your vendor questions, today. The Agents API in beta is US-only for data residency and does not support Zero Data Retention. Read that second one twice. Since August I've been telling operators to ask their agency whether it can produce a session record of what its agents did on their account. This release adds the other half of that question: where does the record live and can it be made not to. An agency running your listing work through this in beta cannot offer you ZDR. That is not a scandal — it's a fact with a date on it that your legal team will eventually ask about, and it is far cheaper to ask in September than in a renewal negotiation.
And the field nobody will cover: whose sandbox. Your agency's agent now runs in OpenAI's environment, its own, or one of nine third parties. That's a new row in your vendor chain that did not exist a week ago, and it sits between the model and your Seller Central session.
What I'd do this week
1. Add one line to the vendor question set. You already ask which model they pin, whether delivery preserves image metadata, and what identity their automation runs under. The new one: which execution environment does your agent work run in, and who hosts it. One email. A shop that has thought about the write path names a sandbox. A shop that replies with a paragraph about responsible AI has told you something too, and that reply is the answer.
2. Ask whether anything touching your account is running on a beta API. Not to forbid it. To know. Beta means US-only residency, no ZDR, and terms that change without a memo — the same category of quiet change I've written about four times this year.
3. Re-price nothing. If you were told orchestration was the hard part and you're paying for it, this is the week to ask what specifically is being maintained now. If you were told the hard part was judgment, nothing about your arrangement changed and you should say so out loud so nobody re-opens it in December.
4. Write down your three concurrency-sensitive jobs. Catalogue-wide attribute fills, bulk image QA, review reading across a long tail — the jobs whose value depends on finishing inside a window. Three-at-a-time changes the arithmetic on all of them, and it's better to know that before someone promises you a two-week catalogue audit.
5. Do not migrate anything before January. Nine weeks from peak is the wrong moment to put a new class of tool on a live catalogue write path. Run it in Q1 against twenty of your own real SKUs, when a wrong answer costs a Tuesday instead of a week of Q4.
What I'd ignore
The agency-extinction cycle and its mirror. Both are content. The thing that moved is that a plumbing layer became free, which is neither a threat nor a defence.
Benchmark comparisons against Anthropic's or Google's agent stacks. Still nothing in any published eval measures whether an agent writes a bullet that won't get your listing suppressed. That remains the only benchmark that bills you when it's wrong, and you still build it from twenty of your own SKUs.
Anyone selling an "agent architecture audit." It is a one-page tool inventory with an execution-environment column added. I have now published that page free three times.
The nine partner sandboxes as a strategic question. It is a strategic question for people who build platforms. For you it is one row in a document, filled in by an email.
The urge to build your own. I built mine because I wanted to understand where it breaks, and what I learned is that it breaks at naming rules, resolution mismatches and the moment a source document gets treated as evidence — none of which the harness fixes for you, and all of which are still yours.
The part I don't enjoy
I spent an evening writing an orchestrator and a longer time writing the sentences that stop it doing something stupid, and on Thursday the first half became a managed product I could have had for free. That is the correct outcome and I'd have made the same trade again, because the evening wasn't the expensive part either — and because the thing that finally made the pipeline useful on its first real run was that it refused to generate three images it didn't have good enough source files for.
No managed harness ships with that refusal. You write it, you own it, and it is the only part of the stack that has ever been worth a retainer.
Ask your agency where their agents run. Not because you'll go and look — because the answer tells you whether anybody over there has thought about what their software is standing on.