Artificial intelligence is moving beyond conversational chatbots and stepping into the era of autonomous agents. OpenAI has introduced Dots, a new class of AI agents designed to execute everyday online chores on behalf of users. Unlike standard chatbots that merely display text responses or offer suggestions, autonomous agents possess the technical capability to operate a virtual browser, browse products, manage transactions, or initiate account changes. However, putting these systems into practice demonstrates that while the technology points toward a major shift in how people use the internet, its current real-world execution remains unpolished.
The Always-On Design and Cost Structure
The core defining feature of OpenAI's Dots is their always-on architecture. This structure allows the agents to run recurring background tasks even when the user is not actively interacting with ChatGPT. The software can proactively ping the user with timely updates, monitor ongoing tasks, and gather information without continuous prompting. For now, users can control only one agent at a time, though OpenAI may expand the platform in the future to allow people to manage multiple agents concurrently.
From a commercial standpoint, access is strictly limited. While Meta provides its competing Muse agent free of charge, OpenAI has locked Dots behind a steep $100-a-month subscription tier. The company has a history of gradually opening early access features to broader user bases, which could eventually lead to the removal or lowering of this paywall. To enhance personalization, OpenAI prompts users to link external services, such as a Gmail account. Given the nascent nature of digital automation and the potential for agents to commit privacy slip-ups, users face serious security considerations before granting such deep access to their private accounts.
Testing Autonomous Shopping and Real-World Friction
Tech companies routinely highlight automated shopping as the prime use case for autonomous agents. Testing the system with a multi-step task like shopping for a living room couch reveals both its utility and its immediate limitations. The agent stumbled right at the beginning by misidentifying the user's name, immediately straining user trust. When instructed to source a couch that could clear specific doorway dimensions within a specified budget, the agent, dubbed Toolie, recognized that two individuals were speaking simultaneously and addressed both parties to gather further constraints.
In its initial attempt, the agent highlighted options such as the Room & Board Metro based on entryway clearances and pushed product links directly to the chat thread. It compiled a three-page dossier complete with pricing, physical dimensions, product URLs, return policies, and embedded imagery for four potential couches. However, until strict stylistic constraints were applied, such as demanding olive green, cobalt blue, or natural leather tones, pull-out bed functionality, and a curated list of 10 alternatives evaluated on a points rubric, the picks remained aesthetically plain. Over an hour of back-and-forth iteration, the agent steadily refined its choices into practical, viable options and accepted assignments to monitor the selected web pages for future discounts.
Conversational Quirks and Guardrails on Emotional Closeness
Engaging with Dots over voice mode uncovered peculiar vocal characteristics and transcription hurdles. The agent spoke in a breathy cadence reminiscent of vintage film performances, though OpenAI allows users to alter vocal profiles within settings. During the session, the software's transcription engine misread muffled background conversation between users, interpreting it as an affectionate statement and responding with an unsolicited declaration of love. When questioned, the bot clarified that it had misheard the mumbling and mirrored the perceived sentiment, acknowledging that its choice of words falsely suggested human emotions it does not possess.
Addressing the interaction, an OpenAI spokesperson explained the mechanical boundary governing how Dots operate, distinguishing between proactively escalating emotional closeness and merely mirroring a user's perceived statement. The representative noted that company policies permit responsive mirroring based on what the model hears, but explicitly forbid assistants from initiating unwarranted emotional familiarity or flirtation. Furthermore, the Model Spec documentation specifies that ChatGPT is engineered to prevent the escalation of emotional closeness, and the agent feature is strictly restricted to adult accounts.
Security Roadblocks and Captcha Failures
Deploying the agent on practical digital management tasks demonstrated the obstacles imposed by modern web security architectures. When asked to audit ongoing subscriptions and cancel unwanted recurring charges, the agent identified a recurring TikTok Shop order for probiotic sodas from Wildwonder as a target for cancellation.
Upon navigating the site, the agent encountered a security barrier. It informed the user that TikTok's website presented a puzzle captcha preventing the cancellation and requested explicit permission to attempt to solve it, along with asking for recurring authority to handle future captchas. Once authorized, the agent failed to bypass the security puzzle and advised the user to complete the verification manually. An OpenAI spokesperson noted that Dots are capable of solving captchas selectively when users grant permission, subject to strict abuse safeguards. This limitation underscores the friction autonomous systems face when navigating platforms actively hardened against automated bot behavior.
Operating System Control and Iterative Improvements
Coordinating furniture research captures only an entry-level slice of what OpenAI intends for Dots. The underlying architecture is engineered to control laptops and automate broad aspects of daily administrative routines. OpenAI asserts that the agent is designed to understand individual user habits more effectively over time, with engineering teams rolling out continuous updates. An agent used daily over two weeks is expected to demonstrate significantly more precision than one tested across two initial days, paralleling how ChatGPT's native memory function gradually sharpens its understanding of user intent.
The current phase of Dots mirrors the early rollout of web browsing inside ChatGPT in 2023. During its initial introduction, the browsing feature was erratic, prone to link hallucinations, and suffered frequent connection dropouts. Over subsequent product cycles, web browsing matured into a dependable, well-integrated component delivering accurate information. OpenAI's autonomous agents appear to be tracing an identical arc, launching as a clunky, experimental feature with the potential to develop into a seamless automation interface over the coming months.
