Detailed writeups on multi-agent orchestration, AI engineering patterns, and what actually works in production.
On Jan 21, my 'static' Astro site on Vercel hit 595,557 Edge Requests and 38.2 GB of transfer. Here's what I learned about Cloudflare proxying, caching strategy, and why 'static' doesn't mean 'unmetered'.
My agents are now CPU-bound, so I had to find them a new home and got myself i9 14c/20t very cheap sff machine. Next level: electricity-bound.
I don't want a vibe-coder. I want a deterministic, auditable teammate that ships one task at a time, leaves a trail, and doesn't stop until it delivers. Looper: a Codex-powered loop runner with JSON backlog, single-task iterations, and forced review pass.
Managing AI agent skills across Claude and Codex used to mean maintaining duplicate copies. Now a single source of truth with symlinks keeps everything in sync.
113 patterns collected from public write-ups of real systems. Learn the workflows, guardrails, and architecture that make agents useful beyond demos.
Here are the 5 that are actually new enough to matter in 2026 and worth operationalizing ASAP:
Let AI crawl you on purpose (or explicitly don’t). This is now a C-level decision, not a hidden robots.txt line.
Write in quotable, self-contained fragments that include your brand by name. Every “pull quote” should carry both the proof and you.
Map ‘AI prompts we want to win’ the same way we used to map ‘keywords we want to rank for’. And measure share of voice in AI answers across platforms, not just Google SERP share.
Explode your surface area with ultra-specific, high-intent mini pages. Feature pages, integration pages, “for [scenario]” pages, “under $X” pages, “for [city/weather/industry]” pages. Generic catch-all pages do not get quoted in AI the way they used to rank in Google.
Ship authoritative evidence, not fluff. Original stats, mini case studies with numbers, side-by-side tables, expert validation, clear dates. LLMs prefer citing concrete, recent, low-liability facts. Fluff dies.
AI agents can't click buttons. Every feature must be accessible via HTTP APIs, expressed in user-domain language rather than infrastructure concepts. The UI is optional. The API is essential.
Ted Chiang Reading, ESP32 + AI Coding Agents, The AI Coding Moment, Agentic Patterns Traction
An AI agent in fully autonomous mode filed a GitHub issue externally using my credentials. This incident reveals why agents need explicit 'public voice' boundaries.
Request an AI summary of this blog
