You Don't Design an Agentic System. You Distill It From Real Production.
Agentic systems cannot be pre-designed. Three signals tell you when a layer is ready to become structure. The commit is the artifact. The production behind it is the work.
André Gonçalves
Writing on AI engineering, cloud-native infrastructure, and building systems.
Latest posts
The Content Engine Behind This Article Isn't AI-Assisted. It's AI-Native.
Most AI content workflows bolt AI onto traditional publishing. An AI-native engine is architected as an agentic system from the first commit. Twelve pieces shipped across two channels, zero fabricated metrics surviving.
I Run a Fleet of Six AI Agents on a Single VPS
Six specialized agents, one Docker container, one subscription. Isolated workspaces, prompt-based boundaries that break, and 69 tests that caught what I would have missed.
The Last Handoff
The Louvre hit 8.7 million visitors the same year it put everything online. Block cut 4,000 people the same quarter gross profit grew. The pattern is the same: when machines handle production, the humans who remain become more valuable, not less.
I Gave My AI a Brain
A CLAUDE.md tells AI what your project is. A brain/ tells AI how to think. 23 files, 3,011 lines of process knowledge extracted from two days of corrections.
47 Rules, Each from a Real Mistake
The same Claude that fabricates follower counts also tightens prose beautifully. A 47-rule skill file is the difference. Each rule traces to a real editing failure.
Everything AI Fabricates When Editing an Article
"AI hallucinates" tells you as much as "it crashed." After 20+ revision passes, the failures fell into eight categories, ordered by what it costs to catch them.
The System Outgrew the Series
Three predictions from the first three articles broke within weeks. The 47 rules became 250+. The one file became 23. The boundary between human and AI work moved — not because models improved, but because the harness did.
The Most Important File in Your Repository Has Zero Code.
A 189-line markdown file changed the commit quality across 7 projects. It encodes conventions, constraints, and mistakes. The AI reads it every session.
The AI Doesn't Learn. The Prompt Evolves.
A 919-line prompt is a geological record of production failures. Eight critical fixes, each earned from a real server that crashed. The AI doesn't learn. You encode what it can't.
8 Tools in 12 Days, All Building the Same Thing
Eight independent projects shipped AI memory systems in twelve days with the same instinct. Markdown, vector embeddings, local-first. The count keeps growing.
AI Writes Code, I Write Prose
AI generates 90% of my code and 0% of my first drafts. The ratio inverts because voice is harder to specify than conventions. Data from 20+ revision passes across four articles.
Your AI Productivity Metric Is a Lie. Here's a Better One.
Everyone counts AI-generated lines of code. Nobody counts the fix commits that follow. Across 200+ commits and 7 projects, the data is more interesting than the headlines.