WRITING INDEX
29 published notes
AI agents / systems / operations
Notes on AI agents, memory, healthcare, and the operating details that decide whether a technical system becomes useful.
Personal AI agents will soon handle real work across our email, calendar, and files. Town is a strong early example. The catch is that the security model is not solved yet, and you can read the open problems directly in the agent's own pre-prompt, because the defense for each one lives there.
On April 8, 2026, Meta released Muse Spark. That same day, I extracted its complete system prompt using 17 short messages. No code, no exploits, no special access. Here is every message I sent, the full extracted prompt, and why this reveals how all AI systems actually work.
In a small office, an intern named Lex sits at a desk covered in today's notes.
Hard-won lessons from Snowflake's Production Agents Summit: the counterintuitive 40% context window rule, observability requirements, and why agents are ultimately just software.
Four common traps doom most AI agent projects before they reach production. A recent industry conference revealed the patterns - and introduced a game-changing metric for measuring real success.
A single statement at a Stanford GSB fireside chat inverted conventional wisdom about AI agent success, revealing that models contribute just 30-40% while framework architecture drives the remaining 60-70%.
Showing 1-6 of 29 articles