Two labs geofenced their best models behind US compliance in a single week: OpenAI's GPT-5.6 preview to US partners only, Anthropic's general model shut down and still offline, its cybersecurity model limited to US critical-infrastructure orgs. The export-control logic is real, and so is the lockout for a builder outside the US. Why I had already moved my work onto a model I run myself, and why the border turned local-first from a preference into a position.
The agentic web breaks the old content-for-traffic-for-ads handshake. I designed a platform to inject ads into agent-read text and pay humans to look, then pressure-tested it against unit economics, fraud, and the law and killed it. Where the durable value actually sits: rails, commerce in the loop, and verification.
A model server can only tell the one client it just answered what that response cost. It can't see the queue, the swap thrash, or the throughput across every caller. The router can, because it owns all of it. How I added queue depth, swap economics, rolling tok/s, and Prometheus metrics to Jano without breaking its passthrough.
A dense 27B at Q8 runs my local chat at 15 tokens a second. Its 35B Mixture-of-Experts sibling runs the same prompts at 70. I ran 50 through both to find what the speed costs, and swapped my daily model. A surprise: the 6-bit MoE matched the 8-bit dense.
Every team building AI inside a company hits the same wall: the model doesn't know who you are. Why vector RAG hits a ceiling on relational data, what the graph turn unlocks, where MCP fits, and what context infrastructure actually solves.
On the same day I released Three-Body Agent, Anthropic shipped Managed Agents. Two solutions to the same problem, one solo dev and one frontier lab, no overlap in audience. What the convergence means.
I moved my whole autonomous coding pipeline onto GitHub Actions in two days, because Anthropic shut the door on the previous setup overnight. Notes on what changed and what I would do differently.
182 PRs merged in 10 days, most of it while I was asleep or with my family. The architecture matters less than the discipline around it. Here is what made it work.
The agent solved the problem and the theory stayed in the prompt. Notes on what gets left behind when the code is clean and the model in your head isn't.
The constraint used to be typing speed. Now it's how well you direct, review, and orchestrate. Notes on what one person can ship with the right agent setup.