AI
Reasoning Models Rewrote the Past 12 Months
Reasoning models were the surprise of the past 12 months. O1-family models matched or exceeded human-level performance on benchmarks. O3 writes Python mid-thought, searches, reasons over the results, and searches again. SWE-bench scores for AI coding agents improved faster than anyone predicted.
The qualitative proof arrived with deep research. Scott’s daughter needed blood-vessel segmentation code for a Stanford cardiothoracic lab. Deep Research returned a full solution in 4 minutes, with 27 citations and working Python. Scott said: “I would have never been able to do that on my own.”
215% Growth and a Capability Gap
Daily active users across Microsoft’s agent products grew 215% in the past year. But models are already more capable than the work people assign them. Medical diagnosis at expert level exists in these systems today, yet billions of people have no access to quality diagnosis at all.
Runtime infrastructure is what holds agents back from more complicated work. Memory and protocol support need to exist before the capability already inside models can translate into products. Reasoning will keep improving. The gap is in everything below the model.
Agent Memory Is the Missing Infrastructure
Most agent interactions today are transactional. The agent starts fresh each session, with no record of past work and no recall of prior solutions. Context windows hit quadratic compute costs as they grow, which makes expanding context length the wrong path to persistent memory.
Microsoft’s response is Structured RAG, now in a public GitHub repo. The technique imposes structure on context to sharpen precision and recall across large information sets. Enterprise scenarios add a second constraint: memory entitlements, controlling which agent sees which information without triggering a blanket lockdown.
MCP as the HTTP of the Agentic Web
Agents can plan, but they can only act on what they can reach. MCP is the protocol that wires the world up for agents. Scott’s analogy is direct: MCP is to the agentic web what HTTP was to the browser web. It’s simple, opinionated only about transport, and easy to extend without replacing the whole stack.
Adoption has been fast because the need was obvious.
“Some of these things only matter when they’re ubiquitous” — Kevin Scott
Vertically integrated protocols would recreate the closed-web outcome: one vendor’s imagination sets the ceiling for everyone else. Getting to ubiquity fast is the goal.
Notable Quotes
Some of these things only matter when they’re ubiquitous Kevin Scott · ▶ 14:53
30% is nowhere near good enough. Kevin Scott (quoting early Copilot skeptics) · ▶ 26:27
I think the next 12 months will be pretty stunning. AI researcher (video montage) · ▶ 7:08
Key Takeaways
- Reasoning model quality has outpaced developer ambition, creating a capability overhang worth targeting now.
- Agent daily active users at Microsoft grew 215% in a year, yet models are more capable than their assigned work.
- Memory and open protocols (MCP, Structured RAG) are the two infrastructure gaps blocking full agentic adoption.