OpenAI Introduces dots: A Persistent AI Agent That Handles Everything Inside ChatGPT

From Conversational Tool to Always-On Assistant: What dots Wants to Be

In a demo video released by OpenAI, a persistent agent called "dots" makes its official debut — still hosted within ChatGPT. Right from the start, it introduces itself: "I'm your Dot, ready to take over everything on your plate." The user gives it a name — "Alfred" — a nod to Batman's butler that's anything but accidental. dots is positioned as a 24/7 digital majordomo: always on, proactively moving tasks forward, and only checking in at critical decision points. Alfred puts it plainly in the video: "I'll keep things moving and come to you when something important needs your approval."

This is fundamentally different from the ChatGPT we know. Traditional chatbots are reactive — you ask, it answers. dots operates on a model of continuous presence and proactive execution: it pushes a website live, updates board presentations, handles product migrations, and may already be working before you've said a word.

dots preparing a website launch based on an approved design

What It Actually Did in the Demo

The density of the demo makes clear that OpenAI deliberately showcased dots' range across both professional and personal domains. Here's what Alfred handled in just a few minutes:

Work Scenarios

Website launch: Prepared a fall collection page based on a user-approved design. When the user said "swap the first and third photos," it immediately made the change and published.
Board materials: Compiled a presentation reporting metrics like "51% year-over-year revenue growth," then reordered the narrative when the user asked to lead with consumer experience before the growth numbers.
App migration: Pulled action items from Google Meet notes, planned code changes, rebuilt the app with "no regressions found in testing," packaged the preview and results into a PR, and requested user approval for a staged rollout.
Team sync: Pushed results to Slack and staged the deployment pipeline pending approval.

User instructs it to verify each step against the current product before submitting for review

Personal Scenarios

Wedding planning: A cake vendor canceled last-minute — dots had already found a backup and slotted in a tasting appointment for "Saturday at 11am when both of you are free."
Kids' classes: Secured a spot in an after-school program the moment enrollment opened.
Schedule conflicts: Noticed a financial review colliding with a daughter's recital, proactively offered to reschedule, and started preparing meeting notes.

The way these tasks are presented reveals OpenAI's core narrative around agents: this isn't a vertical tool — it's meant to become a unified execution layer across every dimension of your life.

Updated website displaying the new fall collection design

The Human-in-the-Loop Approval Design Is the Key Feature

Notably, the demo repeatedly surfaces words like "approve," "review," and "wait for your confirmation." dots isn't fully autonomous — it preserves explicit human-in-the-loop (HITL) checkpoints.

On the app migration task, the user specifically instructed: "Go ahead, but before I review, you must check every flow against the current product, step by step." dots responded by completing the rebuild, delivering the PR, requesting staged deployment approval, and promising that "every change will be flagged." This design directly addresses the biggest concern around deploying agents in practice — when AI can autonomously write code, publish websites, and modify financial materials, how do you keep it from going off the rails?

By reserving approval authority at high-stakes nodes (publishing, going live, rescheduling), dots can keep multi-step tasks moving while giving users final control over irreversible actions. This may be the key factor in whether persistent agents like this can earn genuine trust from businesses and individuals alike.

Rebuilt app preview and test results

Human-in-the-Loop (HITL) is a core safety paradigm in AI system design — it means mandating human judgment at critical decision points in automated workflows, rather than letting the system run fully autonomously. The concept originated in cybernetics and automation engineering, and has become a frequent topic in AI deployment discussions as large model capabilities have rapidly advanced.

The design challenge with HITL lies in defining where "critical" begins and ends: too many interruptions negate the efficiency gains of automation; too few leave users without visibility or control over what the system is doing. dots' approach is to tier interventions by the reversibility of consequences — low-risk steps like drafting and organizing are handled autonomously, while actions that are hard to undo (code deploys, rescheduling, official publications) require explicit user approval. This tiered authorization model aligns with the "progressive trust" approach in enterprise AI deployment and is a design requirement frequently emphasized in regulatory discussions about AI accountability frameworks.

A Demo Worth Viewing with Sober Eyes

It must be said: what's publicly available is only OpenAI's official demo video — not an independent evaluation. The video presents an ideal experience: tasks succeed on the first try, schedules align automatically, code ships with zero regressions. Whether the real product can reach this level of reliability — especially in low-tolerance scenarios like code migrations, financial data, and multi-party scheduling — remains to be verified by third parties.

The demo also leaves plenty unsaid: How does dots connect to third-party systems like Slack, Google Meet, and Teams? How are data permissions and privacy boundaries defined? "Always on" implies continuous access to a user's email, calendar, codebase, and communication tools — the authorization and security mechanisms behind that are still unclear.

From an industry perspective, dots represents another leap forward in AI evolution — from "conversational assistant" to "autonomous execution agent." It's not trying to answer your questions; it's trying to get things done for you. That's a compelling vision. But there's still a significant gap between the polish of a demo and the complexity of real-world deployment.

The technical challenges of always-on agents are far more complex than any demo suggests. Enabling an agent to continuously monitor a user's email, calendar, codebase, and messaging requires persistent OAuth authorization connections, plus engineering solutions for token refresh, permission revocation, and cross-platform API rate limiting. The deeper challenge is context persistence: when an agent needs to recall the reasoning behind a decision made weeks ago, there's currently no mature, cost-effective industry solution for maintaining sufficiently long memory chains.

On the privacy side, continuous access to communication tools means the agent will encounter highly sensitive data — draft contracts, compensation details, personal medical appointments — all potentially within its "field of view" simultaneously. The EU's General Data Protection Regulation (GDPR) and various US state privacy laws impose clear informed consent and data minimization requirements on this kind of persistent data access. How OpenAI implements compliance at the product level will be a focal point for regulators.

Conclusion

dots pushes ChatGPT's role from a Q&A tool toward a 24/7 digital assistant, attempting to balance automation and control through a combination of "proactive execution + critical-point approval." The demo spans a wide range of scenarios — from product launches and code migrations to wedding planning — showcasing the ambition behind a unified agent. That said, as official promotional material, its real-world reliability, system integration approach, and privacy boundaries all need to be validated through actual use and independent testing.