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.
July 2026
Why AI pilots die at handoff
A pilot needs an owner, a quality signal, a runbook and an operating budget before the team can depend on it.
Build vs buy for AI agents
Compare buying and building separately for each layer: model access, orchestration, business rules and data.
AI receptionist vs AI intake agent
Decide whether you need a message or a structured intake record, then check routing, validation and handoff against that requirement.
Claude Code vs Cursor, what one does that the other can't
Compare Claude Code and Cursor by the task: exploring a file, making a scoped change, running tests and reviewing the result. Choose the workflow your team can check.
n8n AI agents in production, what actually breaks
An n8n agent needs more than a model connection. Review budgets, failed tool calls, repeated actions, source updates and prompt changes before handing it to an operations team.
OpenAI Codex CLI, a real operator's guide
Use Codex CLI for a scoped repository task, then inspect the diff and checks. Keep the work isolated and review it before merging.
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
Estimate build cost, model spend, hosting and maintenance separately. A launch budget alone leaves ongoing work unaccounted for.
How to review agent-written pull requests
A tidy diff still needs checks for scope, behavior, APIs, permissions and tests that verify the requested change.
Putting AI coding agents in CI
Give CI agents scoped credentials, isolated jobs and review gates. Check the diff and test results before merging.
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.
MCP server security and auth
Check who can call each tool, on whose behalf and with which permissions. Carry that identity into the audit trail.
Shipping a production MCP server
Tool names, input schemas, results and errors shape how an agent uses an MCP server. Test them before release.
April 2026
Self-hosting n8n at scale
Self-hosting n8n includes operating the database, workers, retention rules and backups. Size and test those parts alongside the workflows.
Choosing an AI agent framework in 2026
Choose an agent framework by the state, restart behavior and handoffs your workflow requires.
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.
February 2026
Context engineering for AI agents
Check what the agent receives on each turn: instructions, task state, retrieved sources and recent history.
The Claude Agent SDK in production
The SDK provides the agent loop. Your application still needs scoped tools, permissions, budgets, retries and tests.
January 2026
What an AI automation agency actually delivers
Before hiring an AI automation agency, agree on the software, checks, documentation and handoff the engagement will deliver.
n8n vs Make vs Zapier for AI automation
Compare n8n, Make and Zapier against your workflow steps, run volume, hosting requirements and ability to investigate failures.