Building multi-agent systems on AWS. Cutting through the hype with working code and real judgment.
About
I build AI infrastructure for a living — multi-agent orchestration, serverless backends, the kind of systems that actually have to work at 3am. I've shipped enough "revolutionary" products to know the difference between genuine capability and marketing copy dressed in a demo.
I'm opinionated about how software gets built and skeptical of anyone who's never debugged something in production. I update my views when I get better data — which is more often than I'd like to admit.
Right now I'm deep in the agentic AI wave, building systems where models coordinate autonomously on real business problems. It's more interesting than the benchmarks suggest, and harder than the demos make it look.
Now
Deep in AWS AgentCore — orchestrating agents that coordinate on infrastructure, email, code, and research. MCP just hit 97M installs. The protocol layer for agentic AI is quietly becoming as important as REST was for the web.
Jack Dorsey cut 40% of Block's staff and cited AI automation. That's not a prediction — it's a press release. The companies being quiet about this are the ones you should watch more carefully.
GPT-5.4, Gemini 3.1, Grok 4.20 — all in March. OpenAI raised $110B at a $730B valuation. At some point the infrastructure layer becomes more defensible than the model layer. We may be close to that inflection point.
OpenAI signed a Pentagon deal days after Trump ordered agencies to drop Anthropic. The geopolitics of AI are moving faster than the policy infrastructure. Nobody has a coherent framework yet — they're all making it up.
Mistral Small 4, Sarvam's 105B sovereign model — the distance between frontier closed models and good open-source is compressing faster than the hyperscalers want to admit. Interesting for anyone building on top of this stack.
Writing
The Model Context Protocol just crossed 97 million installs. When a protocol grows that fast, the early majority is about to show up with bad assumptions and good intentions. Here's what they're going to get wrong.
OpenAI's valuation makes sense if you squint at it right — but the real money is probably in the pipes, not the water. A look at where defensible value actually accumulates in the AI stack.
40% of Block's workforce. Gone. "AI automation" cited in the press release. The companies being polite about this are making the same decision — just slower and with better PR.