TL;DR
The short version
Microsoft CTO Kevin Scott argues the model stopped being the bottleneck. Agents get useful when the world gives them an internet to act on — open protocols, memory, and identity — not when the next model ships.
The winning bet is understanding a real problem better than anyone, then tuning available infrastructure to it. From Every's AI & I, recorded at Microsoft Build.
Source: "Microsoft's Vision for an Internet Made for Agents" — Kevin Scott with Dan Shipper, Every / AI & I (recorded at Microsoft Build 2025).
The model stopped being the bottleneck
For two years the AI conversation was about scaling laws — bigger models, bigger infrastructure, predictable jumps. Scott says that debate is settled and boring now. The models work; they keep improving.
The interesting problem is the opposite of what people expected. Reasoning has run ahead of what products actually do with it. Scott calls it a "capability overhang" — the gap between what the models can already do and what we deliver to users.
That reframes the whole roadmap. If your AI results are underwhelming, the fix probably isn't waiting for a smarter model. It's closing the gap to the capability you already have.
Agents need an internet, not a bigger brain
What closes that gap is plumbing. An agent is only useful if it can take action — use tools, change systems, pull from rich sources. And that requires an ecosystem that looks like the internet.
Scott's analogy is the spine of the whole talk: MCP (the open protocol for connecting agents to tools and data) is to the agentic web what HTTP was to the internet. NLWeb, Microsoft's project for making any site or API agent-ready, is the HTML equivalent — and every NLWeb endpoint is an MCP server by default (VentureBeat).
The lesson for anyone running a product: if you have a website or an API, the work ahead is plumbing it so agents can talk to it. Not a bigger model. A protocol.
Stop shipping your org chart
The sharpest operating line is about internal structure. Scott invokes Conway's law — you ship your org chart — and says his push as CTO is to make every Microsoft system speak one standard protocol to every internal agent.
The failure mode without that: each team wires its agents its own way, and the seams between teams leak straight into the product as complexity. Fragmented teams build fragmented agents.
This is a leadership move, not a technical one. Standardizing the protocol layer is how you stop your organization's shape from becoming your customer's problem.
Bet on the problem, not the infrastructure
On strategy, Scott kills a popular dichotomy. Everyone frames it as verticalized-and-secure versus open-and-innovative. He calls that false. Permissionless systems cut out low-value gatekeeping middle layers, and AI itself can supply real security — so you don't have to trade one for the other.
The most interesting startups right now are not trying to innovate by building some kind of differentiated infrastructure. They're innovating because they think they have an understanding of a problem that someone has that is better than anyone else.
Kevin Scott, CTO of Microsoft
Infrastructure is now cheap to pick up and tune. The durable advantage is a sharper understanding of a real problem. That's also why he expects many agents, not one to rule them all — diversity comes from product insight, not from owning the pipes.
What to do now
Scott's forward call is blunt: the people still saying "the tech isn't ready" — because it was marginally too expensive or marginally not capable — will fall behind, because everything gets cheaper and more capable every year. Betting on failure costs far more than betting on optimism.
And the smallest, most portable piece of advice, from a 41-year programmer who still codes in vi: be curious about the tools, try them, keep what works. The craft survives the change — the people who explore early are the ones who don't regret it.
