Posts
Notes on software, system design, and building with AI.
The postmortem is already written
Forrester predicts AI data-center upgrades will trigger two major multiday cloud outages in 2026, and power is the binding constraint, so I wrote the postmortem now while I can still be wrong in public.
Rate limits weren't built for an agent army
An unattended agent fleet hammering a map API shows what request counts miss. I would check identity and budgets instead.
Three hundred tokens a second
OpenAI now sells speed itself with Ultrafast. I guess at where 300 tokens a second comes from, then argue latency is the next price dial.
The smallest model in your agent stack
Decision models answer one question: what should the agent do next. Jev named the category, AWS and Cloudflare shipped open follow-ons, and I would rebuild my agent loop around one.
