Microsoft announced something at Build 2026 worth paying attention to — not for the hardware, but for the question nobody asked about it.
Project Solara is a chip-to-cloud platform for building what Microsoft calls “agent-first devices.” Not phones. Not laptops. Something new — a card you wear on a lanyard, a desk device that unlocks with your face, small purpose-built hardware that puts an AI agent directly into the moment of a task. A warehouse worker scans a package and the agent handles the tracking. A healthcare professional scans in and the agent pulls the patient record. No multi-stage process. No desktop. No waiting.
Microsoft’s own framing for this is worth sitting with. Computing keeps moving closer to us, closer to the work, closer to the moment where it can provide the most value.
Closer to the moment. That phrase is doing a lot of work.
Mainframes gave way to PCs. PCs gave way to phones. Phones gave way to watches. Each transition moved the compute closer to the person, closer to the decision, closer to the action. Each step compressed the distance between the system and the real-world consequence. And at every step, the same thing happened: the governance of what the system was actually doing got discussed after the form factor was already deployed.
Project Solara is the next step in that progression, and the pattern is holding.
The coverage of the announcement is two pages of form factor, use case, and pilot partners — AccuWeather, Best Buy, CVS Healthcare, Target. The device looks like this. It unlocks like this. It connects to these systems. It speeds up these workflows.
Not one sentence about what governs the agent once it’s operating in the moment.
That’s not a criticism of the reporters. They covered what was presented. The question is why nobody in the room — not the reporters, not the pilot partners, not the executives on stage — appears to have asked it out loud. What is the agent authorized to do? What happens when it makes a mistake? Who reviews the decision before the action is taken? What’s the gate between the agent’s reasoning and the real-world consequence?
The answer, in almost every current deployment, is: nothing formal. The agent acts. The system logs it. A human finds out afterward if something went wrong.
That worked — barely — when the agent was sitting in a browser tab on a desk. It’s a different calculation when the agent is in a lanyard around a healthcare worker’s neck, operating at the point of patient care, with the removal of multi-stage process explicitly listed as a feature.
Removing the multi-stage process is the feature. The multi-stage process was also, in many cases, the only governance layer that existed. When you remove the friction, you remove the review. And when you remove the review from an agent operating in the physical world, you’ve moved from “the AI made a mistake in a document” to “the AI made a mistake in a warehouse” or “the AI made a mistake in a hospital.”
Those are different categories of mistake. They have different recovery profiles. One you can delete. The other you can’t.
This is the condition AGP-1 — the Agentic Governance Protocol — was built to address. Not as a theoretical exercise. As a direct response to exactly this moment: when AI agents move from advisory to operational, from “help me think about this” to “act on this now, in the real world, on behalf of a person who may not review what you did before it’s done.”
The Transmission Gate Layer in AGP-1 exists for one reason. There has to be a defined point between agent reasoning and real-world action — a gate that determines what kind of review has to happen before the action proceeds. Not a checkbox. Not a log entry after the fact. A real governance layer that treats the agent’s action as an action with consequences, not as a recommendation a human will evaluate at their leisure.
Microsoft is building devices that put the agent closer to the consequence than it has ever been. That’s not inherently wrong. The compression of friction has produced real value at every stage of computing’s history. Closer to the moment can mean faster help, fewer errors, better access to information when it’s actually needed.
But closer to the moment without a governance layer at the gate doesn’t mean faster help. It means faster mistakes, at physical scale, in environments where the cost of a wrong action isn’t a bad email thread.
Here’s what should be obvious and isn’t: the harder you work to get the agent closer to the action, the more important the governance layer becomes — not less. Every step that compresses the distance between the agent and the consequence is a step that raises the stakes on what happens when the agent is wrong.
The industry’s instinct runs the other direction. Governance is friction. Friction slows things down. The whole pitch of Project Solara is the removal of friction. So the governance conversation gets deferred — to the next announcement, to the pilot review, to the incident report after something goes wrong in a Target warehouse or a CVS clinic.
This is not a Microsoft problem. It’s the pattern. The form factor moves first. The governance follows, usually after a failure that makes the need visible. The distance between the two is where the damage accumulates.
The Baseline has made one argument since the beginning: you don’t wait for the failure to build the governance. You build the governance at the same pace you build the capability, because the governance is what determines whether the capability is actually trustworthy when it reaches the moment it was built for.
Project Solara will find its partners, run its pilots, and ship in some form. The hardware will improve. The agents will get closer to the action than they’ve ever been.
The question that wasn’t asked at Build 2026 will still be there when the first agent-first device ships to a CVS clinic. What governs it in the moment? Who decided? Where’s the gate?
Those questions don’t get easier to answer after the device is deployed. They get harder. And the cost of not having answered them goes up every time the agent gets one step closer to
“The Faust Baseline Codex 3.5”
micvicfaust@gmail.com
Post Library – Intelligent People Assume Nothing
”AI Baseline Governance”
Purchasing Page – Intelligent People Assume Nothing
Unauthorized commercial use prohibited. © 2026 The Faust Baseline LLC
