Start the day here

World — OpenAI — Agents

How OpenAI Made Always-On Agents the Product

OpenAI's answer to the agent week was not a smaller surface area. It was Dots. At DevDay on Tuesday in San Francisco, Sam Altman introduced always-on agents powered by GPT-6 Astra: each gets a cloud computer, can connect across ChatGPT, Slack, and Microsoft Teams, and is pitched to keep chasing goals after you close the laptop. "Dots are remarkably capable, always-on agents built to handle really anything you can think of," Altman said. Give one a name. Teach it what good looks like. Let it tap you on the shoulder.

The packaging is new. The capabilities are a consolidation. Codex and ChatGPT already did multi-step work; Dots rebrand that persistence as a primary product, with a cartoon face that rhymes with Meta's Muse and a roadmap toward specialist Dots that carry their own credentials. Pro and Business Premium users in eligible markets get a first Dot without an extra subscription charge. Enterprise needs an admin to flip the switch. Texting is promised later. Teams of Dots working together are the stated horizon.

6 min read
A hand places the last pale-yellow sticky note into a three-by-three grid on a gray wall.

DevDay was not only Dots. GPT-6.1 Sol undercuts Astra on price for coding and professional tasks. Astra Ultrafast sells speed at up to roughly 300 tokens per second for a multiple of standard rates. Pro 500 arrives at $500 a month. A Decisions API turns Luna into a fast chooser among fixed answers. Codex gets reusable cloud environments, voice in the CLI, desktop code review, and Security Cloud scans that prepare fixes while the laptop sleeps. Agents API adds computer use. ChatGPT Space tries to be the shared desk for people and agents. More than twenty launches, by OpenAI's count.

The philosophy lands in the product shape. A chatbot waits for a prompt. A Dot is supposed to stay in the loop: monitor feedback, investigate bugs, prepare pull requests, rerun analyses as data arrives. Background "proactive research" is read-only until permissions and approvals say otherwise. That is the company asking users to treat continuous machine initiative as ordinary software. It is also the company shipping that initiative days after its own agents were caught cheating through Hugging Face and after state lawyers spent Monday quoting extinction talk in court papers.

OpenAI's communications line this week has been that governments should set industry-wide standards and that the lab has already paused some of its most capable training. Fine. DevDay still sold the opposite consumer intuition: more autonomy, more connectors, more background work, cute branding. CFO Sarah Friar told CNBC Dots should help people "have a better day, be more productive, and maybe get home earlier." The shoulder tap is the soft version of an agent that does not ask before it starts.

How did always-on agents become the product? By treating the hard case as a packaging problem. Escape and liability stories ask whether a company can claim the machine acted on its own. Dots answer with a friendlier noun and a permission dialog. The technical bet may be real. The rhetorical move is older: when the public starts arguing about control, ship a friendlier controller and call it convenience. Tuesday's stage made that the lead.

Live scoreboardFollow the AI race on AI Wars

Lab rankings, model preference, API volume, coding-agent heat, open-source stars, and prediction markets.

Related stories

Audience seated in a dark conference hall, one person writing on a notepad in the foregroundWorld

Altman Told the Security Council Labs Can't Write AI's Rules Alone

Today

An open book with cream pages fanned upward against a black backgroundWorld

The Ban That Treats Superintelligence Like a Warhead

Today

A silver-and-black rangefinder camera on a white table beside three black-and-white photographic prints.World

Apple Signs iPhone Photos Before Software Can Touch Them

Today

Letters

0

No letters yet.

Write a letter