01 / Field notes
Writing from the middle of shipping.
Notes from building AI systems that actually run. n8n agents, Claude Code, MCP servers, retrieval, evals, the parts nobody writes tweets about. 25 posts and counting.
July 2026
Why AI pilots die at handoff
The pilot worked. Nobody owned it on Monday. That is the whole story, most of the time.
Build vs buy for AI agents
Buy the commodity. Build the part your customers would notice if it disappeared.
AI receptionist vs AI intake agent
Answering the phone is the easy half. The value is in what happens to the information afterward.
Claude Code vs Cursor, what one does that the other can't
Six months of building production apps with both. Where Claude Code wins, where Cursor still wins, and the workflow I actually use day to day.
n8n AI agents in production, what actually breaks
Everyone can wire an OpenAI node to a webhook. Very few can keep it running for six months. Here is what falls over first, and how to build so it doesn't.
OpenAI Codex CLI, a real operator's guide
Codex CLI is not another shell wrapper. It is an agent that edits your repo. Here is how to install it, wire it into a real workflow, and not let it wreck main.
From vibe coding to production
The prototype was real work. It just is not the same work as running it for other people.
June 2026
What AI workflow automation really costs
The model bill is rarely the expensive part. The expensive part is the six months after launch.
How to review agent-written pull requests
The diff looks clean. That is exactly the problem.
Putting AI coding agents in CI
An agent with repo write access and no gate is not automation. It is an incident with a schedule.
AI agent memory, what to persist
Most agent memory systems are a vector store and a hope. Memory is a schema problem first.
A RAG pipeline that survives real documents
Your retrieval is not bad. Your PDF parser gave the embedder a wall of ligatures and page furniture.
May 2026
Prompt caching, and the rest of the bill
Most agent bills are 70 percent repeated context. Caching is the cheapest engineering win available.
Prompt injection is not a prompt problem
You cannot instruct your way out of injection. You architect your way out of it.
MCP server security and auth
An MCP server is a remote code execution surface with good manners. Treat it that way.
Shipping a production MCP server
Writing an MCP server takes an afternoon. Writing one that agents use correctly takes a week of tool design.
April 2026
Self-hosting n8n at scale
Self-hosted n8n is cheap until the executions table hits forty gigabytes. Then it is a database job.
Choosing an AI agent framework in 2026
Framework choice matters far less than people argue about, and in exactly two situations it matters enormously.
March 2026
AI agent observability, what to log
When an agent misbehaves at 2am, you do not need more logs. You need the right six fields.
LLM evals that actually catch regressions
If your eval suite has never blocked a deploy, it is decoration.
February 2026
Context engineering is the whole job now
Your agent is not dumb. It is looking at the wrong 8,000 tokens.
The Claude Agent SDK in production
The SDK makes the first agent easy. Production is a different sport, and almost all of the work happens around the model, not in it.
January 2026
What an AI automation agency actually delivers
The category is two years old and already full of people selling strategy decks. Here is how to tell a builder from a broker.
n8n vs Make vs Zapier for AI automation
All three will demo an AI workflow in ten minutes. Only one of them is still cheap and debuggable at fifty thousand runs a month.