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. 26 posts and counting.
June 2026
AI agent memory, what to persist
Separate scratchpad, task state, durable facts and audit history. Give each a source, an update path and a retention rule.
A RAG pipeline that survives real documents
Before changing the embedding model, inspect the extracted text, chunk boundaries and sources that reach the answer.
May 2026
Prompt caching, and the rest of the bill
Measure repeated input before changing models. Caching, shorter context and task routing each address a different part of the bill.
Prompt injection is not a prompt problem
Treat retrieved content as untrusted. Limit tool capabilities and require approval where a mistaken action would have serious consequences.
March 2026
AI agent observability, what to log
Record the calls, decisions and outcomes needed to investigate a failed run. Use trace IDs to connect the steps.
LLM evals that actually catch regressions
An eval suite needs cases that can reveal a regression and a clear rule for what happens when one fails.