The agent is becoming infrastructure

The easy version of the agent story is that models got better at following instructions. The more useful version is that the surrounding system is becoming the product: context management, tool use, subagents, persistence, and a place where work can safely run.

In its September 2026 announcement, OpenAI described the Agents API as a public beta for building and running cloud agents with the same harness behind Codex. OpenAI says the harness manages context, tools, and coordination while the builder chooses the tools, knowledge, and workflow that make the agent specific.

The wedge is no longer “add a chatbot”

For a small product team, that is a useful constraint. A generic chat surface is easy to copy. A reliable workflow is not. The valuable work moves into permissions, recovery after partial failure, evaluation, source freshness, and the decisions a human must still approve.

The announcement also makes the deployment boundary explicit: OpenAI lists managed sandboxes, customer infrastructure, and partner environments as possible places for agent compute. That matters because the same workflow can have very different privacy, cost, and latency requirements.

The practical takeaway: start with one consequential workflow, define its evidence trail, and make the handoff to a human visible.

What we would measure

Not how many prompts were sent. We would measure completion rate, human takeover rate, time to useful output, error recovery, and whether every important claim can be traced to the source that supported it. Those are boring metrics. They are also the ones that tell you whether an agent is a product or a demo.