The problem

AI infrastructure moves too fast to track by reading and remembering. Every week there’s a new protocol (MCP, A2A, x402), a new platform move (Apple’s App Intents, Cloudflare’s Markdown for Agents), a new benchmark release, a new “this changes everything” post — and almost none of it gets connected to the last one. Most people either drown in the volume and stop reading, or read constantly and retain nothing structural: they can recall the headline six months later but not the pattern it was part of. For a team or a company trying to make build-vs-buy and platform-positioning calls in this space, that’s not a personal inconvenience, it’s a decision-quality problem — you end up reacting to individual announcements instead of recognizing, for example, that three different vendors just made the same “own the orchestration layer” bet in the same month.

The approach

This is a standing research project inside CogitOS, my personal knowledge system, dedicated to one thesis: the web is splitting into a human-facing layer and an agent-facing layer, and the companies that build for the agent client now (the way Uber and Instagram built for mobile-first) will own the next platform era. Concretely, it’s a corpus of roughly 30 processed article/video bundles (Substack posts, YouTube transcripts, links) each broken into an article.md, a summary.md, and a prompts.md, plus a running ideas.md where every source gets reduced to one atomic, attributed insight — not a link dump, a claim with its sourcing kept attached. On top of that sits a synthesis layer: a project README that pulls the individual ideas into one standing narrative (right now: “the agent web is a parallel economic layer, not just a parallel read layer, and Cloudflare/Coinbase/Stripe are already building its payment rails”).

The system has real tooling behind it, not just markdown files: CogitOS exposes capture_idea, save_prompt, and list_projects as callable tools, so when I’m reading or working on something unrelated and hit a relevant insight, I can file it straight into this project’s corpus without stopping to open a folder — it gets routed and merged into the running synthesis automatically. That’s the actual mechanic: continuous, low-friction capture across many sources, feeding one compounding document instead of forty scattered ones.

What I learned

The individual insights turned out to be more durable than I expected — the “middleware trap” framing (you’re structurally exposed if you sit between a model provider and a customer they could serve directly) and the “agent-accessible vs agent-native” distinction (an MCP wrapper over a human API is not the same as infrastructure built for machine speed) both kept reappearing as the right lens for unrelated news weeks later. The bigger lesson was about the capture method itself: the value isn’t in any single summarized article, it’s in forcing every source down to one sourced, atomic claim before it goes in — that discipline is what makes the corpus synthesizable instead of just archived. A pile of full summaries doesn’t compound; a pile of one-sentence claims with their source attached does.

Where this could go

The natural extension is less about this specific topic and more about the pattern: any team tracking a fast-moving external landscape — a competitor set, a regulatory area, a technology shift — has the same problem I had, at higher stakes. An enterprise research or competitive-intelligence function doing this manually in shared docs hits the same failure mode: volume outpaces synthesis, and six months in nobody can point to what changed and why. The same capture-then-synthesize loop, with tool-level filing instead of copy-paste, is the template — the specific subject (agent web vs. human web) is just the first place I applied it. Where it goes next for me is using the synthesized pattern (own infrastructure, not middleware) as an actual filter on my own project decisions, not just a reading list.

Takeaway

Reading a lot of AI news isn’t the same as understanding the shape of the shift it’s part of — this project is my attempt to force the second one to actually happen.