OpenAI Just Opened the Agents API. Marketing Still Builds Workflows by Hand.

Yesterday OpenAI did the thing every agency Slack has been arguing about for a year. It put the Codex harness on a public API.

Not a toy playground. Not another chat window. A managed agent loop: sessions, orchestration, context compaction, recovery. You bring the tools and the job. They run the plumbing.

If you still brief AI like a junior copywriter, this week just made you late.

What actually shipped

The Agents API went public beta on September 10. One call can stand up a production agent. OpenAI hosts the harness. You pick the environment: their sandbox, your stack, or a partner box.

Early numbers are the only part I care about. SafetyKit cut cost per case by 60% after moving review work onto it. Hypha dropped agent failures 86% by splitting the loop from the sandbox. Cirridae cut latency 4x. No extra API tax. You pay tokens and tools.

• LAYER 1 — The loop is now infrastructure. Stop building orchestration as a side project.

• LAYER 2 — Long-running work is the product. Briefs that die after one prompt are dead weight.

• LAYER 3 — Multi-agent is no longer a slide. Subagents and MCP tools are in the same pipe.

Marketing and AI: agents will not save a lazy funnel

Forrester already said the quiet part. Nine in ten US agencies use generative AI. Half are touching agentic AI. Most of that spend is going into speed, not into better work. That is how every shop starts sounding the same.

The marketing use of this API is not "write 40 ads." It is a specialist that stays on a brief for hours. Pull last month's winners. Check brand kit drift. Draft variants. Score them against the offer. Hand a human a shortlist, not a landfill.

Same week, the other half of the job did not move: generative engine optimization. Answer engines still decide if your brand gets named. An agent that ships volume into a void is just a faster way to be invisible. Pair the agent with a GEO layer or you are decorating a page nobody asks for.

• Lock one agent to one job. Creative scoring. Competitor page watch. Weekly GEO citation audit. Not all three.

• Give it the brand kit, the offer, and the last five winners. Starve it of vibes.

• Measure brand match and answer-engine mentions, not how many files landed in Drive.

Do this before Friday

Pick one painful weekly task. Write the runbook in plain English. Point an Agents API session at it with two tools and a kill switch. If the output is not something you would send a client, kill the session. Do not add a second agent.

We are not collecting models at Dzine Prodigy. We are renting loops that finish work. This week the rent got cheaper. Use it.