· Software
OpenAI Is Now Selling the Agent Plumbing Most Engineering Teams Were About to Build
Buying the orchestration layer instead of building it is usually the right call, and OpenAI's new Agents API is honest proof of both halves of that trade.
OpenAI opened its new Agents API to every developer in public beta on September 10, and the pitch is blunt: stop building the orchestration layer yourself. OpenAI's own announcement describes putting the same managed harness that runs Codex, its coding agent, behind a single API call. Instead of writing the code that tracks conversation history, compacts a growing context window, recovers from a crash mid-task, and coordinates subagents, you send OpenAI the work and its servers run the loop. Pricing is simple: no separate charge for the harness itself, just standard rates for the model, the tools it calls, and the sandbox minutes it uses, according to OpenAI's developer documentation.
My position: this is a correct and overdue buy decision for a lot of engineering teams, and OpenAI is more honest about the tradeoff than most vendors selling infrastructure. The convenience is real. So is the dependency you take on to get it, and which side of that trade actually favors you depends on whether the orchestration logic was ever going to be your product's advantage.
What building this yourself actually costs
Steelman the case for buying it first, because the arithmetic is not close. Session state that survives a restart, a context compaction strategy that does not silently drop the instruction the agent needs three turns from now, retry logic that recovers from a failed tool call without duplicating work, a sandbox where the agent can run code safely: none of that is exotic engineering, but all of it is slow, easy to get subtly wrong, and invisible to a customer when it works. Assume a mid-level engineer, at roughly $145,000 a year in salary per the Lightcast data Fortune cited on a related hiring trend, spends six weeks getting that plumbing production-ready. That is roughly $16,700 in salary alone, before the ongoing maintenance every time a new failure mode shows up in production. The Agents API charges nothing extra for the equivalent, forever, on top of tokens you were already going to pay for. OpenAI's launch partners back that up with real numbers: SafetyKit reported a 60 percent reduction in cost per case, and Hypha saw an 86 percent drop in failed agent responses, according to OpenAI's own launch case studies. Treat those as a ceiling, since a vendor picks its best examples for a launch post, not an average. But the direction is not in question. Debugging your own retry logic is not where most companies' AI product actually earns its keep.
What you give up to get it
Here is the honest complication. The same coverage that reported those results also named the real tradeoff plainly: adopting the Agents API means locking your orchestration layer to OpenAI's specific implementation of it. That matters less if your product's edge is the tools you built and the workflow you designed, and the orchestration underneath was never going to differentiate you from a competitor running the identical loop. It matters more if you expected to swap models or providers as prices moved, because you are no longer just choosing a model each time, you are depending on however OpenAI decides context compaction and crash recovery should work this quarter, with no visibility into the logic and no fallback if it changes in a way that breaks your specific workflow. That is a real cost, and it is the same cost every "just buy the platform" decision carries. The one mitigating factor: OpenAI's own open-source Codex codebase, which the managed API is built on, stays available, so a team that outgrows the hosted version has a real starting point for building its own rather than starting from nothing.
The actual decision
Ask one honest question before adopting this: is the value my company delivers the orchestration itself, or the judgment built into what the agent does with it? For nearly every business outside a company literally selling agent infrastructure, the answer is the second one, and buying the plumbing to spend engineering time on the judgment layer instead is the right call. That is the same build-versus-buy audit MojoSoftware runs before recommending a client build anything custom: assume buying the commodity part is correct until the specific reason it is not shows up in your own workload, not a hypothetical one. If you are trying to make that call on your own AI product right now, that conversation is worth having.
Sources
References used in this article. Links also appear alongside the relevant claims.
Tell us what's on your mind.
You don't need a polished brief to reach out. A two-line email about what's bugging you is plenty; we'll tell you straight if we're the right fit, and what we'd tackle first.
We'll scope the work around your workflow, goals, and timeline before quoting anything, so you know what's included before committing.