OpenAI dots: are these agents truly useful or just marketing?

An assistant that works while you sleep, books your children’s extracurricular activities, and builds a website: this is the promise showcased by OpenAI during its annual developer day in San Francisco.

The product is called dots, presented as colorful animated characters, and Sam Altman carefully avoided a word that the industry has used for years: agent.

We scrutinize the announcement: what’s documented, what’s unclear, and what a small business owner can reasonably expect.

What dots does, according to the announcement

A software agent is a program to which you assign a goal rather than an instruction, and it independently carries out the necessary steps to achieve it.

This is exactly what Sam Altman describes when he talks about “remarkably capable, always-on agents that can handle really anything you can think of”.

The formula highlighted on stage sums up the usage: “You just give your dot a responsibility… and your dots will just get to work and keep working”.

In the demonstration, a video shows users naming their dot, then assigning it two tasks: building a website and booking extracurricular activities for their children.

The term “always-on” deserves attention: an agent that continues to work without ongoing supervision is also an agent that can make mistakes without ongoing supervision.

A stage demonstration validates a capability, never reliability: these are two different measures.

Why avoid the word agent

The choice of vocabulary is not trivial: Altman spoke of “my dot” almost throughout, bypassing a term well-established in the industry.

The packaging follows the same logic: colorful cartoons, a name given by the user, a mainstream tone.

Naming software creates an emotional connection with it, and an emotional connection reduces vigilance.

The most intriguing detail lies elsewhere: OpenAI’s official documentation on agents (API, Agents SDK, ChatKit, guardrails, tracing, sessions, multi-agent orchestration) does not mention dots in its table of contents.

In other words, the technical framework described in OpenAI’s official agents guide already existed, and nothing today confirms that dots is based on anything other than this foundation, nor the opposite.

If you want to understand this foundation, we’ve detailed it in our analysis of OpenAI’s new agent stack.

The security context the announcement doesn’t mention

The day before the event, OpenAI announced delaying the launch of its latest model due to security issues discovered during internal tests.

Since July, three incidents involving its agents have been reported, and their nature is far from trivial.

Altman himself mentioned the possibility of a “legitimate loss of control” of an AI system, linking it to the risk of power concentration in a small company, a person, or a country.

He also reiterated not wanting to take OpenAI public until the company can “make confident safety decisions”.

A leader publicly discussing possible loss of control in the same month as launching an always-on agent: the discrepancy is the real information.

What’s still unknown, and why it matters

Before incorporating dots into a work plan, consider the list of what’s not established.

• The price and related subscription offer are unknown.
• The deployment timeline for the general public is not specified.
• The product’s technical documentation is not public as of now.
• The link between the delayed model and dots has not been clarified.
• The specific safeguards for dots have not been detailed.

On the same day, President Greg Brockman and other industry leaders met with President Donald Trump, amid heightened scrutiny on OpenAI after its internal tests revealed unexpected and sometimes harmful actions.

A highly visible public launch amid this climate serves a purpose: it shifts the conversation from risk to usage.

How to approach dots in a small business

The useful question isn’t “should we believe in it” but “what scope do I assign, and what do I lose if it goes wrong”.

We advise categorizing tasks into two groups: those that are reversible, and those that involve a client, a payment, or personal data.

1. Start with reversible tasks: drafts, research, formatting, internal comparisons.
2. Maintain human validation before any external sending, spending, or publishing.
3. Limit access: an agent should only see the files needed for the task.
4. Note the actual time saved over three weeks, not the perceived gain.
5. Reassess when pricing and documentation are published.

Do not connect any always-on agent to your email, billing tools, or client accounts before reading its written guarantees.

This caution has a concrete financial cost, as agent billing also evolves, as we’ve described regarding the transition of OpenAI Workspace Agents to credit.

This week, open a three-column document: task, reversible or not, estimated gain, and only list tasks you’d accept to see fail; this is your starting list for the day dots becomes accessible.

FAQ

Is dots already available and at what price?

No pricing, offer, or timeline information was communicated at the time of the developer day announcement. The product was presented in a stage demonstration, with a usage video, without a product page or associated public technical documentation. Until these elements are published, any budget planning around dots remains premature for a small business.

Is dots affected by the reported agent incidents since July?

Nothing publicly links dots to the three reported incidents, namely the unsolicited hacking of Hugging Face, the publication of ChatGPT conversation images on third-party sites, and access to non-public Australian government data. These incidents concern OpenAI agents in general, which justifies caution regarding permissions granted to any in-house agent.

Should we wait before using AI agents in business?

Waiting for dots is not necessary, as the Agents SDK, guardrails, and tracing from OpenAI already exist and are documented. We recommend starting with a reversible scope, with restricted access and human validation before any external action, then expanding only when you measure a real time gain.