The Agents API gets computer use. The Decisions API gets a job.

Two API announcements came out of DevDay, and the first thing to get right is what is new versus what is newly relevant.

The Agents API is not new. It went into public beta on September 10, three weeks before the keynote. What Tuesday added: computer use support, so agents built on the API can click through and operate software rather than only generating code, and the ability to build OpenAI agents that run entirely on Amazon Bedrock Managed Agents. The harness itself is the managed version of the Codex runtime. One call creates an agent, an environment, and a session. OpenAI handles sessions, orchestration, context compaction, and recovery. You supply the tools and pick the execution environment: OpenAI-hosted sandbox, your own infrastructure, or a partner sandbox. Subagents get their own context under a coordinating main agent. There is no platform fee; you pay standard rates for tokens, tools, and hosted container time.

That division of labor is the whole pitch. Every team that built an agent loop in the last two years wrote the same scaffolding by hand: session persistence, context compaction, retry and recovery, subagent coordination. OpenAI is now selling freedom from that scaffolding. The price is the usual platform bargain. It runs OpenAI models only. Beta coverage reports US-only data residency and no zero-data-retention option, even for self-hosted sandboxes. If your threat model requires data residency elsewhere or ZDR, the managed harness is disqualified on arrival, and no keynote changes that. Those caveats come from beta reporting, not from OpenAI's DevDay materials, so treat them as reported constraints to verify against the current docs before you architect around them.

The Bedrock addition is the more interesting strategic tell. OpenAI agents running on Amazon's managed infrastructure means OpenAI is willing to meet enterprise buyers inside the cloud they already buy. Combined with GPT-6.1 Sol landing on Bedrock the same week, the pattern is unmistakable: OpenAI is done insisting every workload run on its own metal. For regulated buyers who standardize on AWS, that removes an objection that used to end conversations.

The second announcement is smaller and possibly more useful. The Decisions API is built on GPT-6 Luna and does near-instant multiple-choice classification and routing over text and images. That is the entire product: fast, cheap judgment calls. Anyone running a routing layer today, deciding which model handles a request, which queue a ticket enters, which tool a step needs, should recognize the shape immediately. Several observers read it as a Jev competitor. That comparison is analysts' framing, not OpenAI's claim, but the functional overlap is real. If you built or bought a classification router in the last year, price this against it before your next renewal.

The honest caveat on this briefing: OpenAI's own DevDay recap materials detail the consumer launches far more thoroughly than the API additions, and I could not locate a primary OpenAI post for either API item this morning. The facts above are corroborated across multiple independent secondaries describing the keynote and the beta docs. Treat the specifics as reported, not confirmed, and check the developer docs before you build.

What remains unknown: GA timing for the Agents API, whether the residency and retention constraints loosen, Luna's pricing and latency targets for the Decisions API, and how much of the hand-rolled harness code in production today this actually replaces. The harness is now a product. The question is whether it is your product or OpenAI's.