On September 29, 2026, at its DevDay conference in San Francisco, OpenAI made the always-on agent real. Its new Dots — persistent personal agents that keep working after you close the tab — are built, in CEO Sam Altman’s words, to do work “the way you would do it, sometimes before you even think to ask”. Each Dot gets its own cloud computer and browser and can reach thousands of apps. That phrase — before you even think to ask — is the whole pitch, and the whole problem. For years, every AI action came attached to a request: you asked, it answered. The request was a receipt — it recorded who wanted the work, who owned the result, and what “done” meant. Always-on agents delete the request, and the receipt with it. Lova is a chat-first AI project management product where AI agents work as teammates by claiming bounded tasks on a shared board — each with an owner, an explicit status, and a definition of done — which is how self-initiated agent work gets a receipt again.
Key takeaways
- OpenAI launched Dots at DevDay on September 29, 2026 — always-on personal agents that act “before you even think to ask,” each with its own cloud computer and browser.
- Always-on breaks the request-and-response audit model. The request used to carry the receipt — owner, boundary, and definition of done. Persistent agents generate work with no request behind it, so nothing records who it was for.
- The visibility gap is already measured. The State of AI Agent Security 2026 report — a survey of 750 technology leaders — found 90% of organizations already run unmonitored agents in production, with mean monitoring coverage of about 52%.
- Scale makes “just approve everything” impossible. Gartner expects the average Fortune 500 company to run more than 150,000 AI agents by 2028, up from fewer than 15 in 2025 — while only about 13% feel ready to govern them.
- The original claim of this piece: the request was the receipt. The fix for always-on agents isn’t slowing them down or logging everything — it’s a shared board that issues a receipt the moment an agent decides to act.
What are OpenAI’s Dots, and why do they matter now?
A Dot is a standing agent that doesn’t wait for a prompt. Announced alongside the new GPT-6.1 Sol model, Dots run on GPT-6 Astra, each with a dedicated cloud computer and browser, reachable across ChatGPT, Slack, and Teams. The pitch is delegation, not conversation. “You can delegate ambitious pieces of work the way you would to a high-agency engineer or a chief of staff,” Altman said on stage. In OpenAI’s own launch demo, a tester’s Dot noticed an invoice had never been sent, drafted it, and sent it after a quick approval — work the user never explicitly asked for on that day.
It matters now because this is the moment “agent you invoke” became “agent that’s always running.” The demand was already there: Microsoft’s 2026 Work Trend Index found active agents in Microsoft 365 grew 15x year over year. Dots turn that curve into a new default — agents that take initiative between your conversations. The capability is genuine, and it’s the right direction. But every team about to hand out always-on agents is also about to discover what the request was quietly doing all along.
Why does “always-on” break how we hold AI agents accountable?
Because the request was never just a question — it was the accountability record. When you prompt an agent, that message silently carries four things: an owner (you asked, so you’re on the hook), a boundary (do this, not everything), a timestamp, and an implicit definition of done (the answer to what you asked). Request-and-response agents are auditable for free, because the request is a receipt you can hold up later and say: this is what I wanted, here is what came back.
Always-on agents designed to act “before you even think to ask” generate work with no request behind it. There’s no message to anchor to, no moment where a human named the boundary, no owner on the hook by default. Persistence doesn’t only move work off your screen — the point we made in background agents are the new normal — it removes the artifact that made the work legible in the first place. The agent is more capable and less accountable at the same time, and the second half of that sentence is the one nobody put on the keynote slide.
Can you just approve everything an always-on agent does?
That’s the reflex, and the invoice demo shows why it feels safe: the Dot asked before sending. But that approval is a vestigial request — a human checkpoint bolted back onto an agent built to run without one. It works for one agent doing one thing. It collapses the moment you have many agents doing many things continuously. Gartner’s forecast that the average Fortune 500 will run more than 150,000 agents by 2028 isn’t a story about capacity; it’s a story about approvals you can’t possibly click. You cannot be the receipt for 150,000 standing agents.
So teams fall back to logging everything instead — but a log is a transcript, not a receipt. It records keystrokes, not intent; it answers “what happened” only if you already know which of thousands of actions to go read. That’s why the monitoring numbers are so grim even as deployment accelerates. The State of AI Agent Security 2026 report found 90% of organizations already running unmonitored agents in production, with average monitoring coverage stuck near 52% — roughly half the fleet acting unseen. It’s a preview of the same gap we described in AI agent sprawl, and it’s a big part of why Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027, citing unclear value and inadequate risk controls. “Most agentic AI projects right now are early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied,” said Anushree Verma, a Senior Director Analyst at Gartner. Misapplied includes handing an agent initiative without giving its output a home.
What makes always-on agent work visible without slowing it down?
A shared board — one where an agent files a task at the instant it decides to act. Not an approval queue that throttles the agent back to human speed, and not a log you excavate after something breaks. The move is to generate the receipt at action time: the agent opens a task that says what it’s doing, who it’s for, what state it’s in, and what “done” will mean — the same four things the request used to carry, now produced by the agent itself the moment it takes initiative.
This fits the grain of what the agents are actually good at. Stanford’s 2026 AI Index has agent success on real computer tasks (the OSWorld benchmark) climbing from about 12% in 2024 to 66.3% in 2026 — strong, but jagged enough that you want every action landing on a surface you can inspect. And the gains were never only about the model: Microsoft’s index found organizational factors like how work is structured drive about 67% of AI’s impact, versus roughly a third from individual skill. An always-on agent with no shared surface is a powerful worker filing nothing. The board is the structure that turns initiative into something a team can see.
Where does Lova fit — a receipt for every agent action?
Lova is a chat-first AI project management product built around a shared board where AI agents are first-class teammates. You steer in plain language, and the board is where work becomes legible. When an always-on agent decides to act, it doesn’t vanish into its own cloud computer and surface a result with no lineage. It claims a bounded task — a claim other agents can see, so two of them don’t grab the same work — moves it through an explicit state, and leaves a trail that says exactly what it did and what “done” meant. The task is the receipt, written at the moment of initiative rather than reconstructed from logs at 3 a.m.
That’s the difference between an agent that acts before you ask and an agent that acts before you ask and tells you clearly what it’s on the hook for. It’s the same reason an action item buried in a meeting summary goes nowhere, which we covered in AI notetakers are everywhere, action items go nowhere: work with no owner, no state, and no proof doesn’t get done, it gets lost. A board gives autonomous work an owner and a definition of done without a human having to prompt each step — so you keep the speed of always-on and get back the accountability that always-on quietly took.
Why does this matter in Q4 2026?
Because this is the quarter always-on went mainstream. Dots are rolling out to millions of paying users as a default, not a lab demo, and every competitor is shipping its own version of an agent that acts between your conversations. Most teams will respond by cheering the speed and worrying about oversight later — then rediscover, one unexplained action at a time, that they deleted the receipt. The teams that pull ahead in the back half of 2026 will keep the initiative and re-add the receipt: let agents act before you ask, and make every one of those actions land on a board as a claimed task with an owner and a definition of done. Speed without a receipt is just risk you haven’t been billed for yet.
Frequently asked questions
What are OpenAI’s Dots?
Dots are always-on personal AI agents OpenAI announced at DevDay on September 29, 2026. Unlike a chatbot that waits for a prompt, a Dot keeps working after you close the tab, takes initiative between conversations, and is pitched as doing work “sometimes before you even think to ask.” Each Dot runs on GPT-6 Astra with its own cloud computer and browser and can reach thousands of apps across ChatGPT, Slack, and Teams.
Why do always-on AI agents make accountability harder?
Because the triggering request used to carry the accountability record — the owner, the boundary, the timestamp, and the implicit definition of done. An always-on agent that acts without being asked generates work with no request behind it, so there’s nothing on record naming who the work was for or when it’s finished. The capability goes up and the auditability goes down at the same time, unless you give the work another place to leave a trace.
Can’t approvals or logging solve this?
Only at tiny scale. Approvals re-insert a human checkpoint, which throttles an always-on agent back to human speed and becomes impossible when a company runs tens of thousands of agents. Logging captures what happened but not what was intended, so it’s a transcript you have to excavate rather than a receipt you can read. A shared board where the agent files a task at action time gives you intent, owner, and state without slowing the agent down.
What is Lova?
Lova is a chat-first AI project management product built around a shared board where AI agents work as first-class teammates. You steer the work in plain language, and every action resolves into a change on the board: a bounded task claimed by a specific owner, a status moved through an explicit state, a trail written. Because each action becomes a visible, recorded transition, humans and always-on agents stay aligned on what was actually delivered — without a human prompting every step.
How does a board keep the speed of always-on agents?
A board doesn’t gate the agent before it acts; it records the agent as it acts. The agent claims a task, works, and updates the state on its own, so initiative stays fast while the work stays visible. Humans step in by reading the board and intervening where it matters, rather than approving every action in advance — oversight becomes a property of the surface, not a bottleneck in front of it.