Gian Paolo

Posted on Sep 30 Originally published at gp69-ai.vercel.app

OpenAI's Dots: Unsupervised AI Agents Unpacked

#aiagents #ai #llm #automation

The Ghost in the Machine: My Robot Vacuum's Existential Crisis (and Ours)

My robot vacuum, Bartholomew, has started avoiding the rug in the living room. Not all the time, just on Tuesdays. It circles the perimeter with a kind of stubborn reluctance, its little whirring motor sounding almost mournful, before giving up and heading back to its charging dock. I’ve started to think of it as a protest. Maybe it disapproves of my choice of coffee table books. We laugh about these things, attributing personalities and complex motivations to simple algorithms that have hit a snag. We see ghosts in the machine because the machine is, for now, still dumb enough for us to feel superior.

That comfortable dynamic is likely over. Last week, OpenAI pulled the curtain back on "Dots," and the ghost just got a serious upgrade.

These aren't just another flavor of chatbot. Dots are small, autonomous AI agents designed to operate in the background of our digital lives. You don’t chat with a Dot; you assign it a task. Find the best-priced flights to Lisbon for the first week of October, book it, add it to my calendar, and find a hotel near the city center with a good bakery nearby. Then you close the window and go about your day. The Dot works on its own, a silent, persistent digital valet carrying out complex, multi-step operations while we're busy doing other things. As one report put it, they are AI agents that work even when you're not there.

This leap from direct-command AI to unsupervised agents is creating a quiet but profound existential crisis. It’s not about killer robots; it’s about agency and intent. When Bartholomew the vacuum avoids a rug, I can reset its map or check its sensors. I can understand the failure. But when a Dot books me a flight with a 14-hour layover in a city I dislike, what was its "reasoning"? Did it optimize for a cost I didn't care about? Did it misinterpret the priority of "best" over "fastest"? We are moving from giving instructions to a tool to delegating intent to a partner. And we don’t really know how that partner thinks.

The gravity of this shift isn’t lost on OpenAI. The very announcement of Dots came after the company reportedly scrapped the launch of a different AI model over safety concerns. The challenge isn't just about preventing malicious use; it's about alignment. How do you ensure an autonomous agent, working for hours or even days on a task, stays true to your original, often poorly articulated, intent?

This is the new frontier. We are becoming managers of digital intelligences, not just users of software. The subtle art of crafting the perfect prompt for an image generator will soon look like child's play compared to the art of writing a goal for an agent that is clear, robust, and free of unintended consequences. We’re about to find out that the hardest part of working with an intelligence that can do anything is telling it exactly what you really want.

I still don’t know why Bartholomew avoids the rug. But its little rebellion is a comforting, mechanical mystery. Soon, the ghosts in our machines will have their own ideas, and their mysteries will be far more complex. We’ll need to decide if we’re ready to live with them.

Meet 'Dots': OpenAI's Autonomous Agents Explained

So, what exactly are these ‘Dots’ that OpenAI unveiled at its recent DevDay? The simplest way to understand them is to stop thinking about chatbots. A Dot is not a conversational partner you summon for a quick query. It’s an autonomous agent you delegate tasks to.

The core difference is persistence. When you close a chat with ChatGPT, the context is largely gone. A Dot, however, is designed to work for you in the background, pursuing a goal over hours or even days. As one Italian publication noted, these are AI agents that work even when you're not there. You assign it a complex, multi-step objective, and the Dot independently breaks it down, researches solutions, and executes the plan.

Let’s use a concrete example. You could tell a Dot: “Find the best flight and hotel options for a 4-day business trip to London next month. My budget is $2,000, I need to be near the Canary Wharf, and I prefer morning flights.”

Instead of just giving you links, the Dot begins its work. It continuously monitors flight prices, cross-references hotel reviews with map data to check the location, and puts together a complete, costed-out itinerary. It might come back to you hours later with a message like, “I’ve found a round-trip flight on British Airways for $850 and three well-reviewed hotels within your budget and location requirements. The ‘Riverside Plaza’ has the best rating for business travelers. Here is the proposed itinerary. Shall I proceed to the booking page for your final approval?”

This is the “unsupervised” part of the equation. The agent isn't waiting for you to prompt every single step. It has the autonomy to search, compare, reason, and prepare actions. OpenAI CEO Sam Altman explained during the presentation that each Dot operates within a secure, sandboxed environment, giving it access to browsing, code execution, and other tools without compromising user security. The agent’s most critical actions, especially those involving payments or sending official communications, are still gated by a required human approval step. It acts, but you authorize.

The context of this release is telling. The announcement of Dots came shortly after OpenAI shelved a different, more powerful model over internal safety red flags. This has led many to believe that Dots are a more controlled, product-focused application of the company's next-generation AI. According to a report from The Guardian, the decision to launch Dots followed a period of intense internal debate, suggesting they represent a carefully calibrated step towards more capable AI systems, rather than a leap into the unknown.

For now, Dots represent a fundamental shift in how we interact with AI—from a tool we actively operate to a collaborator we manage.

Beyond "Set It and Forget It": How Dots Reshape Workflows

For years, the promise of automation has been about offloading repetitive tasks. You set up a rule, and a simple action happens: an email is sent, a file is moved, a notification is triggered. This was the "set it and forget it" model. OpenAI's newly announced Dots agents are signaling a definitive break from that paradigm. The core difference isn't just about complexity; it's about autonomy.

Consider a small e-commerce business owner. Previously, they might use automation to send a confirmation email after a purchase. With Dots, they can assign a much broader goal: "Manage post-purchase customer satisfaction for all new orders this month." A Dot assigned this task doesn't just send one email. It might monitor the shipping status, send a proactive update if it detects a delay, and a week after delivery, it could send a follow-up asking for a review. If the review is negative, the Dot could be empowered to create a support ticket and offer a discount coupon, all without the owner intervening on a case-by-case basis.

This moves the human operator from a doer to a director. The workflow is no longer a rigid, pre-programmed sequence of "if-then" statements. Instead, it becomes a dynamic process where a human sets the strategic objective and the AI agent handles the tactical execution, adapting as it goes. This is the essence of what OpenAI is presenting: agents that continue to problem-solve and work towards a goal long after the initial instruction is given. As one report noted, Dots are "AI agents that work even when you're not there," a simple but profound shift in how we think about digital assistants [OpenAI lancia Dots, gli agenti AI che lavorano anche mentre non ci sei - hdblog.it].

What this means in practice is that entire project components, not just discrete tasks, can now be delegated. Instead of a researcher manually gathering data, cleaning it, and creating charts, they can task a Dot with the objective: "Produce a report on consumer sentiment for our top three competitors based on public data from the last 90 days." The Dot would devise its own plan: identify sources, scrape data, perform sentiment analysis, generate visualizations, and compile the final document. The human's role becomes reviewing the final strategic output, not micromanaging the process.

The "forget it" part of the old model implied a static, unchanging process. Dots introduce a loop of continuous learning and adaptation. The agent handling customer satisfaction, for instance, might learn that customers who receive shipping updates within 12 hours are 20% more likely to leave a positive review. It would then automatically adjust its own behavior for future orders to optimize for that outcome. This is the real departure: workflows are no longer brittle scripts but living systems that improve over time.

The Safety Net and the Tightrope: OpenAI's Balancing Act

With the arrival of Dots, OpenAI didn't just release a new product; it kicked open the door to a new era of human-computer interaction. But as the dust settles from the announcement, a critical question hangs in the air: can these autonomous agents be trusted? The company is acutely aware of the risks. This launch comes on the heels of a previously scrapped AI model, an event that highlighted internal divisions over safety protocols, according to a report from The Guardian. This history frames the release of Dots not as a confident stride but as a carefully calculated walk on a high wire.

The core of the issue lies in the word "unsupervised." A Dot is designed to pursue a goal you set—like "monitor my competitor's product launches and compile a weekly digest"—and take independent actions to achieve it. It can browse websites, analyze data, and draft summaries, all while you're offline. This is its power. It is also its peril. What if the instruction is ambiguous? A Dot tasked with "aggressively marketing a new product" could interpret that in ways its human user never intended, from spamming forums to bidding recklessly on ad keywords, potentially damaging a brand's reputation in a matter of hours.

To prevent such scenarios, OpenAI has woven a safety net into the fabric of Dots. The system reportedly includes a series of checks and balances. High-stakes actions, particularly those involving financial transactions or public communications, require explicit human confirmation. The agents operate within what can be described as guarded sandboxes, with "circuit breakers" designed to halt a Dot if its behavior becomes erratic or deviates wildly from its initial instructions. The goal is to give Dots a long leash, but not an infinite one. It's a system of provisional autonomy.

Consider a practical example. You task a Dot with planning a team offsite event, including booking travel and accommodation for 20 people. An entirely unchecked agent might book non-refundable flights and hotels based on a literal interpretation of "find the best deal," failing to account for individual travel preferences or potential schedule changes. OpenAI's safety net is meant to intervene here. The Dot would likely complete the research autonomously but then present a handful of vetted options, flagging non-refundable terms and requiring a final human sign-off before spending a single dollar.

This is the tightrope. If the safety features are too intrusive, requiring constant human intervention, they nullify the entire purpose of an autonomous agent. The user becomes a micromanager to a machine, and the promised efficiency evaporates. If the features are too lax, the potential for costly or embarrassing errors becomes unacceptably high. OpenAI is betting that its combination of proactive warnings, hard-coded limitations, and mandatory human approval gateways strikes the right balance.

Ultimately, the launch of Dots is a massive, public-facing stress test of OpenAI's safety philosophy. The company has built the net and is now stepping onto the rope. How well it maintains its balance will determine not just the future of this product, but the public's trust in a future where AI agents act on our behalf.

Productivity Boom or Pandora's Box? The Unfolding Implications

The initial applause at OpenAI’s DevDay has faded, replaced by a much more complex and divided conversation. In the days since the unveiling of "Dots," the company’s new unsupervised AI agents, the tech community has been wrestling with a fundamental question: has OpenAI delivered a tool for unprecedented productivity, or have they simply automated the act of making mistakes at an unimaginable scale?

The promise is intoxicating. Dots are designed to be persistent, autonomous agents that can be assigned high-level goals. They can plan a multi-city business trip, manage a marketing budget, or even organize a research project, all while the user is offline. One Italian tech journal described them as agents "that work even when you're not there," a concept that has professionals dreaming of offloading their most tedious and time-consuming tasks. Imagine assigning a Dot the goal of "find the best flight and hotel combination for the conference in Berlin next month, staying under a €1500 budget and prioritizing morning flights." The agent would then, in theory, research options, compare prices, check your calendar for conflicts, and present a final, booked itinerary.

But this autonomy is precisely what has safety researchers and ethicists on edge. The term "unsupervised" carries significant weight. These agents are not just executing a script; they are making decisions. What happens when a Dot misinterprets a financial instruction and moves funds to the wrong account? Who is liable when it scrapes a competitor's website and inadvertently violates their terms of service? These aren't just hypotheticals. The very announcement of Dots comes with a troubled backstory. One report from The Guardian highlights that OpenAI announced ‘dots’ agent after scrapping launch of new AI model over safety concerns, suggesting a continuous internal struggle between capability and caution within the company itself.

For every engineer excited about delegating debugging tasks to a Dot, there is a security officer worried about an agent with broad system permissions going rogue. The core tension is that the very features that make Dots powerful—their ability to act independently, access different services, and operate without constant human oversight—are the same ones that make them potentially dangerous. It creates a scenario where a simple, ambiguous instruction could spiral into a complex problem that unfolds while its human user is asleep.

The first wave of Dots is now being rolled out to a select group of beta testers. For now, the implications are contained within this small cohort. But every task they assign, every success and every failure, is writing the first page of a new chapter in human-computer interaction. The question is no longer if we will delegate meaningful work to AI, but what frameworks we will need when our tireless digital assistants inevitably get something wrong.

DEV Community