It Hosts Exactly the Layer Everyone Writes Badly
Anyone who has built a long-running agent has written the same batch of code: where to store session state, which stretch of context to cut when the window fills, how to retry a failed tool call, how to resume after the process dies, how to merge parallel subtasks. None of it is business logic, none of it differentiates a product, and it eats a great deal of time at wildly uneven quality. The Agents API takes that layer over. The most immediately useful piece is context compaction: as a session nears its context limit, earlier content is compacted automatically while preserving what the agent needs to continue — and the root agent and each subagent compact independently. That means a task can span multiple context windows without you writing a compaction strategy. Tool search loads tool definitions only as needed, saving tokens while preserving the model's cache. Programmatic tool calling lets calls run in parallel or chained, filtering results in code before bringing only the relevant parts back into context instead of dumping raw returns in. It is not the same thing as the Agents SDK, and that is easy to conflate: the SDK is an open-source framework you deploy yourself, while the Agents API is a managed service where OpenAI runs the harness.
Two Limits Will Decide Whether Some Teams Can Use It at All
The hardest ones first: it is **United States only**, and it is **not eligible for zero data retention** — note that this holds even if you swap in a sandbox on your own infrastructure. For teams with data-residency or compliance requirements, the second is close to disqualifying: ZDR is a hard clause in plenty of enterprise contracts. Settling that before moving a critical path over is cheaper than discovering it after the pilot. Every call also carries the header `OpenAI-Beta: agents=v1`, a reminder of what stage this is at.
What You Get in Exchange
The early-adopter numbers OpenAI lists: Ciridae cut latency to a quarter, SafetyKit cut cost 60%, Hypha reduced failed responses by 86%. These are customer-reported across different workloads — don't read them as what you would get. The more useful calculation is a different one. A hosted harness means your agent's orchestration layer becomes OpenAI's implementation. The upside is a few engineer-months and a pile of edge cases you no longer own. The cost is that this layer is no longer yours to control — which compaction strategy applies, which context gets dropped, you can neither see nor tune. For a product whose core advantage is orchestration quality, that is a trade worth thinking hard about; for a team using agents as internal tooling, it is close to free. Sandboxes leave room: OpenAI-hosted, your own, or Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel — the execution environment stays on your side, and what gets hosted is the control plane.
via: OpenAI, "Introducing the Agents API", OpenAI developer docs: Agents API overview, MarkTechPost technical write-up