SiteMax CMS

[{"h":"CONFIRMED: OpenAI turns the harness into a product","p":"OpenAI opened Agents API in public beta on September 10, 2026. The news is not a new model, but an operational layer that lets developers create cloud agents with the same harness Codex uses to coordinate context, tools, and subagents. In a single call, developers specify the task, model, tools, and execution environment. The agent race is gaining a more important front: who controls the infrastructure that keeps a model working when the conversation becomes long and the workflow becomes complex?"},{"h":"The agent can leave the laptop","p":"The platform offers an OpenAI managed sandbox, while also allowing customers to choose their own infrastructure or partner environments such as Cloudflare, Modal, Vercel, E2B, and Oracle. This flexibility addresses different requirements for networking, storage, CPU, GPU, and memory. In OpenAI’s sandbox, an agent can run code, work with files, and produce artifacts, with configurable packages, skills, and plugins. In practice, OpenAI is trying to shorten the distance between an agent demo and a service that must survive long sessions."},{"h":"Context, tools, and subagents come together","p":"Agents API automatically compacts context as a session approaches its limit, so teams do not have to rebuild memory management from scratch. Tool search loads only the definitions that are needed, reducing tokens and preserving cache. Programmatic tool calling lets agents run operations in parallel and filter results before returning them to the model. The platform also supports MCP, custom functions, web search, and subagents working in parallel, with each subagent keeping its own context."},{"h":"Codex becomes an architecture reference","p":"OpenAI says the product is powered by Codex’s open source harness. That does not mean the entire operation is open, because OpenAI still manages the service, sandboxes, and platform evolution. It does mean developers can inspect the coordination layer and better understand how model calls, tools, and context are orchestrated. The move also turns decisions that were previously hidden inside proprietary infrastructure into a product surface that can be studied, versioned, and compared."},{"h":"The price is simple, but the bill is not","p":"During the public beta, OpenAI says there are no additional fees for Agents API. Customers pay for the tokens and tools their agents use. The sentence is attractive, but it does not equal a predictable cost per task. Compaction, searches, parallel calls, subagents, code execution, storage, and the chosen environment can increase consumption. For a small or midsize company, the metric to measure is the cost of a verified delivery, not only the model’s price per million tokens."},{"h":"Our editorial read","p":"This launch matters because it moves agent competition from the prompt to the entire system. OpenAI is selling a ready path for persistent context, recovery, tools, and coordination, precisely the layers that can consume months of engineering. Our verdict is to monitor and test it on a narrow workflow, without replacing an existing harness by decree. Public beta lowers the cost of experimentation, but it does not yet prove enough reliability, isolation, observability, or rollback for critical operations. The real advantage appears only when the agent completes useful work with less intervention and controlled cost."},{"h":"Sources","p":"OpenAI, Agents API announcement, September 10, 2026: https://openai.com/index/introducing-the-agents-api | OpenAI Developers, Agents API overview: https://developers.openai.com/api/docs/guides/agents-api/overview | OpenAI, open Codex harness repository: https://github.com/openai/codex"}]