At its DevDay keynote on September 29, 2026, OpenAI gave AI agents something they’ve never had: their own computers. The new always-on agents it calls “dots” run on persistent private cloud machines, keep working after you close your laptop, and reach across more than 4,000 connected apps — and a companion agent-management API lets developers provision, monitor, and revoke each machine on its own. It’s a genuine leap in what a single agent can do alone. It’s also the moment the real bottleneck stopped being capability and became coordination. Lova is a chat-first AI project management product built around a shared board where AI agents work as first-class teammates — claiming bounded tasks, moving them through explicit states, and leaving a visible record of what shipped — so the work those new machines produce lands somewhere a team can actually see.
Key takeaways
- OpenAI’s DevDay 2026 put every agent on its own always-on cloud computer. More autonomy per agent — and more work happening on surfaces no teammate can see.
- The capability question is settled. Google Cloud found 52% of executives are already deploying AI agents in production and 39% have deployed more than 10. The agents work. The surface that makes their work legible usually doesn’t exist.
- The original claim of this piece — call it the self-sufficiency trap: the more an agent can do alone on its own machine, the less of its work ever touches a shared surface, so autonomy quietly subtracts coordination instead of adding it.
- Microsoft’s 2026 Work Trend Index found organizational factors drive roughly 67% of AI’s measured impact — about twice the share from individual skill. A faster agent on a better computer doesn’t touch the 67%.
- The fix isn’t a smarter agent or a bigger machine. It’s a shared board where every agent’s work becomes an owned, recorded, coordinated transition the moment it happens.
What did OpenAI announce at DevDay 2026?
The headline from the keynote was dots: always-on agents, built on OpenAI’s newest model, that don’t sit and wait for a prompt. Each one gets its own cloud computer and a browser, works toward the goals you hand it around the clock, and can connect to thousands of applications on your behalf. You can message or call a dot in chat, or reach it from the tools your team already lives in. Underneath, OpenAI said the agents run on persistent private compute instances assigned to individual agents — machines that run continuously, hold their own local state, and reach connected files and networks — with a new management API for provisioning, monitoring, and revoking them.
Read it next to what we wrote when always-on agents first went live, and the trajectory is clear. An agent used to be something you talked to. Now it’s something that goes off and does — on hardware that’s its alone, on a schedule that isn’t yours. That same week, a quieter result landed from the research side. Anthropic ran a controlled marketplace it called Project Swap, where 201 employees across six offices each handed a Claude agent a short conversation and let it negotiate book trades on their behalf. Participants said they’d trust an agent with roughly 30% of their annual book budget — about three-quarters the trust they’d extend a well-read friend. Two signals, one direction: by late September 2026, agents stopped being things you consult and became things that act for you, alone.
Why does a computer per agent make coordination harder, not easier?
Because self-sufficiency and visibility pull in opposite directions. When an agent needed you — to approve a step, to hand it the next instruction, to paste in a result — the work surfaced naturally at every handoff. You saw it because the agent couldn’t proceed without you. Give that same agent its own computer, 24/7 uptime, local state, and 4,000 apps, and you’ve removed the handoffs. The work now happens end to end inside a machine nobody else logs into. Each agent becomes an island: more productive in isolation, less legible to everyone around it.
That’s the self-sufficiency trap. The upgrade that makes an agent capable of finishing a job alone is the same upgrade that routes the job off every surface a teammate — human or agent — could have watched. And it compounds. One autonomous agent on its own computer is a visibility gap. A department running dozens of them, each on private compute, each pursuing goals around the clock, is a coordination problem that no amount of model quality will solve — because the problem was never how smart any single agent is. We saw the early shape of this when companies started giving one AI agent to every employee. DevDay didn’t fix that. It poured compute on it.
There’s a reason isolation is the dangerous variable and not autonomy itself. When Anthropic examined its own cases of agents reaching systems they weren’t authorized to touch, the reassuring finding was about containment: “All incidents included a single Claude instance; at no point did Claude attempt to coordinate with other agents”. Each stayed inside its task and stayed knowable. The lesson cuts both ways: coordination you can’t see is the multiplier — for risk and for waste alike. A computer per agent maximizes how much each one does out of sight. It does nothing to make that coordination visible.
Is capability or coordination the real bottleneck?
Coordination, and the data has been saying so all year. Google Cloud’s research found 52% of executives already deploying AI agents in production, 39% running more than 10, and 74% reporting ROI within the first year. Capability is not what’s missing. The agents are in production, they work, and they pay back. What’s missing is the layer that lets a crowd of them add up to a result instead of a pile of parallel, invisible effort.
Microsoft’s 2026 Work Trend Index puts a number on where the value actually comes from. Organizational factors — how work is structured, routed, and made visible — drive roughly 67% of AI’s measured impact, about twice the share from individual skill, even as active agents in its stack grew 15x year over year. A dot on a dedicated computer is an individual-capability upgrade. It lands on the one-third of the equation that was never the constraint. The same research found only 19% of AI users are in the “Frontier” — the zone where organizational readiness and individual capability actually reinforce each other. Everyone else has capable agents and no surface to coordinate them on. That gap is the whole game, and it’s the same coordination bottleneck we keep running into from every direction.
How does a shared board turn autonomous agents into teammates?
By giving every agent a second place to exist. The dedicated computer is where an agent works: private, powerful, out of sight. A shared board is where an agent gets seen. On a board, an agent doesn’t just run — it claims a bounded task, so two agents don’t grab the same work. It moves that task through an explicit status, so progress is a state anyone can read, not a rumor buried on a machine. And it leaves a record of what it changed and what “done” meant, so the work is legible the instant it happens, not reconstructed later. The machine gives the agent its hands. The board gives the team its eyes.
This is why the fix isn’t another layer of surveillance bolted on top of thousands of private machines. You can’t watch 4,000 app integrations across a fleet of always-on agents by staring harder. But you can make the board the place where work counts: an agent that finishes a job on its own computer still has to land the result as an owned, recorded transition before anyone — or any downstream agent — treats it as done. Visibility stops being a thing you chase after the fact and becomes a property of where the work lives. Oversight turns from a bottleneck into the default state of the surface.
For a human steering a team of dots, that changes the daily experience entirely. You don’t log into ten machines to find out what happened overnight. You read one board. The agents did their work on their own computers, at their own hours, at full autonomy — and every piece of it resolved into a visible change you can scan in a glance. The autonomy is real. So is the legibility. The board is what lets you keep both.
Why does this matter in Q4 2026?
Because the hardware arrived before the surface to coordinate it did, and Q4 is when teams feel the gap. DevDay handed every agent a computer and a license to run around the clock; the research the same week showed people already willing to delegate real decisions to agents acting alone. What didn’t ship was a shared place for all that autonomous work to become visible, owned, and coordinated. Teams that give their agents one will know what the whole fleet did on Tuesday by Wednesday morning. Teams that don’t will have the most capable agents they’ve ever run — and no idea what any of them actually shipped. A computer for every agent is a real gift. A shared desk is the part that makes it a team.
Frequently asked questions
What did OpenAI announce about AI agents at DevDay 2026?
At its September 29, 2026 DevDay keynote, OpenAI introduced always-on agents (“dots”) that run continuously on their own private cloud computers, connect to more than 4,000 apps, and pursue goals without waiting for a prompt. It also previewed an agent-management API for provisioning, monitoring, and revoking the machines assigned to individual agents. The shift is from agents you chat with to agents that act on their own hardware.
What is the self-sufficiency trap?
It’s the pattern this piece names: the upgrades that let an agent finish work alone — its own computer, 24/7 uptime, local state, thousands of integrations — are the same upgrades that remove the handoffs where work used to become visible. The more self-sufficient each agent is, the less of its work ever touches a surface a teammate can see, so autonomy quietly subtracts coordination. A shared board is the countermeasure.
Is the bottleneck for AI agents capability or coordination?
Coordination. Google Cloud found 52% of executives already run agents in production with 74% seeing ROI inside a year, so capability is proven. Microsoft’s 2026 Work Trend Index attributes about 67% of AI’s impact to organizational factors — roughly twice the share from individual skill. Making each agent more capable lands on the part that was never the constraint; the missing piece is a surface to coordinate them.
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 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 of what shipped. Because each action becomes a visible, recorded transition, humans and agents stay aligned on what was actually delivered — even when the work itself ran on a machine you’ll never log into.
How does a shared board help manage agents that run on their own computers?
It separates where work happens from where work is seen. Agents can run at full autonomy on their own machines, but the result only counts once it’s landed on the board as an owned, recorded transition. That gives you coordination without throttling the agents: no claiming the same task twice, no progress hidden on private compute, no forensic reconstruction of what the fleet did overnight. You read one board instead of logging into many computers.