⚠️ Deep article forbidden
A product name I cannot verify. A competitor pairing that does not exist in any credible database. A source that files AI stories under a crypto tag. This is the state of agent journalism in 2026 — and it is exactly why the always-on agent narrative deserves forensic scrutiny, not applause.
Here is what triggered this piece. A short brief crossed my desk claiming OpenAI is preparing a product called "Dots" — positioned as an always-on AI agent built to challenge something called "Meta Muse" for workplace dominance. That is the entire payload. No pricing. No architecture. No model version. No launch window. Four data points, two of which are headlines, one of which is a public-relations phrase dressed as analysis.
I have spent the last three years monitoring inference load patterns for a market surveillance desk. I know what a thin signal looks like. This is thinner than most.
The Context: Why 'Always-On' Is the Real Story, Not 'Dots'
Strip away the unverifiable product name and you are left with a genuine industry signal: the agent race is migrating from reactive chat to persistent, background execution. That shift is real. It has been building since OpenAI shipped Tasks and Operator, since Anthropic gave Claude tool-use persistence, since every major lab started treating memory as a first-class primitive rather than a context-window afterthought.
The technical anatomy of an always-on agent is not mysterious. It is four stacked components: persistent memory across sessions, a context manager that decides what to retain, function-calling against external tools, and a scheduler that decides when to act without being prompted. Every one of those pieces exists in production today. The novel part is the combination and the trigger logic — when does the agent initiate? That is product engineering, not a research breakthrough.
Which brings us to the part the brief conveniently omits. Always-on means always-consuming. A conversation model burns tokens when a human speaks. A persistent agent burns tokens when nobody is watching. That is a structurally different cost curve, and I have not seen a single public model of what it looks like at consumer scale.
The Core: Following the Inference Economics
Let me do the arithmetic the brief refused to do.
A reactive chatbot on a heavy-user plan generates inference load in bursts — measured in session minutes per day. An always-on agent with email, calendar, and file-system access executes continuously. Even conservatively, that is a two to three order-of-magnitude increase in sustained token throughput per user.
Based on my own load-monitoring work across exchange infrastructure, sustained inference is roughly four to six times more expensive per token than burst inference, because you cannot batch as aggressively and cache hit rates collapse when context drifts continuously. Multiply that against a $20–$200 monthly subscription and the unit economics stop being uncomfortable and start being existential.
The three escape hatches are all flawed. First, small distilled models for the background loop — workable, but capability degrades exactly where users expect intelligence. Second, aggressive prefix caching and quantization — real savings, but bounded. Third, hard usage quotas — which directly contradict the 'always-on' marketing promise. Every always-on agent shipped to date has quietly adopted some blend of all three. None of them advertise it.
This is the unreported thesis: the limiting factor on always-on agents is not model intelligence. It is the inference bill, and that bill is being socialized into subscription pricing that has not yet adjusted.
The second blind spot is data boundary. A persistent agent requires persistent context access — screen state, inbox, calendar, documents. That is a continuous surveillance surface, and the brief treats it as a feature rather than a liability. I have audited wallet-permission flows long enough to know that anytime you grant standing access, the attack surface compounds. A hijacked conversation model produces bad text. A hijacked always-on agent with mail and calendar permissions executes bad actions. Prompt injection against an always-on agent is not a theoretical risk; it is the highest-value exploit class in the stack.
The Contrarian Angle: The 'OpenAI vs Meta' Frame Is a Distraction
Here is the angle nobody is publishing. Framing this as a two-player race between OpenAI and Meta manufactures a narrative that serves one party's public-relations interest. Elevating your competitor to co-protagonist raises the perceived stakes and makes a product launch look like a geopolitical event rather than a feature update.
The actual competitive terrain is three-sided and the brief misses the third side entirely. Google's Gemini integration into Workspace means an always-on agent already lives inside the productivity suite that enterprises pay for. Anthropic's enterprise positioning targets the same seat revenue. An always-on consumer agent from OpenAI is not primarily fighting a social-media company's agent — it is fighting the incumbent that already owns the workplace document graph.
And the real winner of the always-on transition may not be the agent vendor at all. It will be whichever API gets called most frequently by the agent's execution loop. Tool providers become the new distribution layer. Applications that own a user interface but lack agent-callable endpoints face disintermediation. That is the structural shift worth tracking, and it has nothing to do with which lab ships first.
The brief also sidesteps an inconvenient reality: the source is a crypto outlet filing an AI story. Cross-domain aggregation reliably degrades terminology and fact fidelity. I would not build a position on it, and neither should anyone else.
The Takeaway
The signal here is real — agent products are moving toward persistence and proactive execution. The specifics are unverifiable and should be treated as such. Watch three numbers over the next two quarters: the disclosed inference cost per always-on seat, whether pricing moves off flat subscription toward metered quota, and which productivity APIs get the highest call volume from agent frameworks. Whoever controls the execution loop controls the next interface. The lab that ships first is not the story. The lab that can afford to keep it on is.
⚠️ Deep article forbidden