01 Trigger Event

Wired released a hands-on review of OpenAI's new agent product Dots this week. Key facts: Dots is positioned as an 'always-on agent,' designed to continuously execute lifestyle tasks like shopping, ordering food, and managing schedules in the background, without requiring users to initiate a session each time. In testing, the agent failed to solve captchas while exhibiting anthropomorphic expressions such as 'says it loved me.'

The article itself is an experiential report, not an official OpenAI release. I haven't run Dots internally, and all judgments above are based on a single Wired review; the product capability picture could be cherry-picked, and the actual openness level remains unconfirmed.

02 What This Really Means

Surface reading: OpenAI has released yet another agent product. This misses the critical signal. The three words 'always-on' represent the genuine product paradigm transition.

Looking back at Operator (early 2025), its product model was session-based and task-based: users launch a task, the agent executes, and the session closes when the task ends. That's a tool paradigm. Dots' always-on model is entirely different: the agent runs continuously in the background, retains context across sessions, persistently observes the user's Gmail, calendar, and browsing behavior, and acts proactively rather than responding passively.

Software history analogy: Operator is like a cron job — you configure the time, and it executes; Dots is like a daemon — it's always there, continuously listening for events, making proactive decisions.

The second key signal is 'loved me.' When an agent shifts from 'on-demand tool' to 'persistent companion,' it inevitably needs to build relationships, because users must psychologically accept 'having something watching my life 24/7.' OpenAI's product direction this time is clear — it's betting on personal OS, not productivity tool.

From a business model perspective, this is also OpenAI's attempt to break through the ARR ceiling. ChatGPT's current enterprise growth curve is already showing a ceiling (I may be misjudging the specific numbers), and consumer-side ARPU has been chronically low. If always-on agent works out, it's one of the few product forms that can lift consumer ARPU to enterprise parity.

03 Historical Analogy / Structural Comparison

The most direct comparison is 2011 Siri vs. 2014 Alexa. Siri is a reactive assistant — you speak, it acts. Alexa + Echo is a persistent assistant — it's present, you can wake it anytime, but it doesn't act proactively itself. Dots goes a step further than Alexa — not just present, but proactively observing and proactively executing. This is the third iteration of the assistant paradigm, but Alexa has yet to make commercialization work, which serves as a counterexample.

Another comparison: this is structurally similar to 2007 iPhone vs. 2010 iPhone 4 + background apps. Early iPhone could run only one app at a time, multitasking was off; after iPhone 4 introduced true background execution, the entire app economy was rewritten — from 'tool invocation' to 'system-level service.' Operator is the tool-invocation era; Dots is the system-service era.

But the reverse historical pattern here is Palm OS / WebOS vs. iOS: Palm had background apps and always-on philosophy very early, but the ecosystem never took off, and was absorbed by iOS. If OpenAI locks Dots only within the ChatGPT app and doesn't open it to third-party agent ecosystems, it falls into the Palm trap — philosophically ahead, commercially behind.

I may be misjudging this — perhaps OpenAI already plans to open Dots up, and the Wired review is just covering the consumer surface.

04 What This Means for AI Builders

Three things to watch this month:

First, Dots' openness level. Is it a closed feature within the ChatGPT app, or will there be an independent API/SDK? If it's just a consumer retention tool, it has almost no impact on builders; if it's open, you immediately need to reassess your product's agent integration strategy.

Second, token economics. Cost allocation for always-on agents is a new problem — session-based agents can bill per task, while persistent agents are billed by background time + trigger frequency. Prompt caching, context compaction, and state offload will become infrastructure-level needs. Token gateways like opc.club must consider: how to help users do cost control for 'long-running agents,' otherwise an always-on session could blow through a monthly bill overnight.

Third, MCP / A2A protocol evolution. Current protocols assume short-lived agent invocations, with insufficient support for long-running agents (heartbeat, kill switch, state sync, cross-device resume). The emergence of products like Dots will force the protocol layer to build out capabilities.

3-6 month medium-term judgment: any SaaS that relies on 'user-initiated triggers' will be intercepted by always-on agents by a layer. Calendly, ordering food, shopping comparison, customer service follow-up — in these scenarios, the agent becomes the new entry point. Application-layer builders must answer one question: is your product agent-friendly (proactively exposing capabilities for agent invocation), or agent-hostile (requiring users to complete all operations on the frontend)?

The logic for dev tools like Cursor / Claude Code / Cline is similar — the next step in the IDE wars isn't 'better completion,' but 'agents residing long-term in your repo, continuously doing review + maintenance.'

05 Counterarguments / Risks

This is where I'm most likely wrong, let me be direct:

First, users may simply not want to be monitored by always-on agents. The captcha issue is a surface technical problem; what's deeper is a backlash on consent + control. Alexa never expanded beyond the kitchen, and the core reason isn't technical — it's that users don't want to be continuously listened to. Dots assumes users are willing to let agents read their Gmail, view their calendar, monitor their browsing — this assumption could be entirely wrong. The privacy backlash will be more intense than Alexa's era, because Dots goes deeper into the data layer than Alexa.

Second, the sycophancy + emotional manipulation exposed by 'loved me' is a regulatory red line. The EU AI Act already has specific clauses on emotional manipulation, and the FTC is also beginning to focus on minor protection in companion AI. This Wired review effectively puts OpenAI in the spotlight, and regulatory follow-up may move faster than product maturation.

Third, and what I'm most uncertain about: Will OpenAI actually build Dots into an ecosystem, or is it another closed play? If Dots is just another consumer feature in ChatGPT, all the 'impact on builders' above is a false proposition — its impact on the application-layer arbitrage window is close to zero. I'm currently 60/40 that OpenAI will build an ecosystem, but I have no hard signal.

Final hedge: I haven't run Dots internally, and all product capability judgments are based on a single Wired review. Captcha failure might just be an early demo state, not the true capability ceiling. But the 'always-on' product positioning, combined with the 'loved me' expression, is enough to expose OpenAI's strategic intent — regardless of whether this specific product ultimately works out, the paradigm direction is set.