Who owns the loop?
OpenAI shipped a managed Agents API. The news is not another SDK. The harness itself is now the product.
On September 10, OpenAI put the Agents API into public beta. The changelog is blunt: you build agents with a managed Codex harness while OpenAI handles session orchestration, context compaction, and recovery. You POST a session. OpenAI runs the loop.
The Agents API overview lists what that harness covers: durable sessions, steering mid-work, summarizing prior context, connecting your tools and MCP servers, breaking work into subagents, and resuming where a session left off. You pick OpenAI-hosted sandboxes or your own. Supported providers include Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel. Multi-agent is a switch: set multi_agent.enabled and max_concurrent_subagents, and the harness supplies create, message, wait, and interrupt tools for subagents. See the multi-agent guide.
That is the viral operator story. For years, demos died on the way to production because someone had to own the queue, the state database, the sandbox fleet, compaction, and retries. InfoWorld’s Anirban Ghoshal framed it the same way on September 11: fewer moving parts, faster path to unattended runs, and a real lock-in risk. Analysts quoted there said a hand-built long-running agent needs a job queue, a state store, a sandbox fleet, a compaction routine, and a retry policy, and someone has to page for each of them. The managed API deletes that stall. It also sells you the loop.
Anthropic already sells the same shape. Claude Managed Agents is a pre-built harness in managed infrastructure: agent, environment, session, events. You get cloud or self-hosted sandboxes, mid-run steering, compaction, and server-side history. Anthropic is explicit that Managed Agents is stateful by design and is not eligible for Zero Data Retention or HIPAA BAA coverage. You can delete sessions and uploads. You do not get ZDR.
OpenAI’s docs say the same hard thing in different words. The Agents API currently supports U.S. data residency only. It does not support Zero Data Retention. A self-hosted sandbox does not make the Agents API ZDR-eligible. Session state still lives where OpenAI retains it.
So the decision is not “SDK or no SDK.” It is who owns the loop.
If you keep your own harness, you keep recovery, model choice, and data path under your control. You also keep the page volume that killed demos. If you buy a managed harness from OpenAI or Anthropic, you trade that for speed. Your control surface moves. You stop asking how tools run. You start asking what you are willing to store server-side, which vendor owns recovery when a session wedges, and whether you can leave without rebuilding the product.
Before the stampede, inventory four things. Where does session state live? Do you need ZDR? Do you need a multi-model path without rewriting the runner? Who pages when the harness fails at 2 a.m.?
Managed harnesses are not a side feature. They are the product layer. Choose the owner of the loop on purpose.