Agents, taken apart.
Past the six parts: how the loop actually spends tokens, when one agent should become several, MCP under the hood, and how to tell if it's getting better. The engineering, not the analogy.
Take the lesson →Context engineering.
Prompting is context management. What to load into the window, what to keep out, retrieval versus memory, and how to test a prompt like it's code.
Take the lesson →Build your stack on purpose.
Stop collecting tools. A stack chosen by role and cost-per-outcome, switching cost and lock-in, and the moment a subscription should become an API call.
Take the lesson →Automations that don't break quietly.
The failure mode isn't a crash, it's a wrong thing done confidently at 3am. Scoping, blast radius, idempotency, the human gate, and watching the thing you built.
Take the lesson →Ship, price, and prove it.
The numbers under a thing that pays: what it costs to deliver with AI in the loop, what to charge, margin math, and the two-week test that tells you if it's real.
Take the lesson →