Architecture
One source of truth: how the site builds itself from the catalog
We had our services stored in the database and written out again on the website. Every change had to be made twice, and one side was always stale.
Blog
What we learn,
written down.
Craft notes on websites, software, AI automation and marketing. What works and what does not.
Context engineering
Agent memory: the problem is not remembering, it is forgetting
An agent that remembers everything is slow, expensive and confused. Designing what gets discarded turned out to be harder than designing what gets kept.
Fundamentals
The fundamentals AI does not replace (and that now matter more)
Idempotence, migrations, soft deletes, failing fast. Decades-old concepts that turned critical exactly when the one writing the code is not always human.
Best practices
Writing rules for agents: every line of our AGENTS.md cost us an incident
A rules file does not get written in one sitting. It gets written after each thing that breaks, and that is exactly why it works.
Architecture
Why our agents never touch the database
Giving an agent direct database access looks faster. It is, right up until the day a creative agent runs an UPDATE without a WHERE.
Workflows
When to use a workflow and when to let the agent think
The most expensive decision in an AI system is not which model you pick. It is which parts you keep rigid and which parts you let reason.
Orchestration
Durable execution: agents that survive the server going down
A process that takes three days cannot live in the memory of a process that restarts on every deploy. Here is how we solved long-running work.
Tooling
Fewer tools, better names: the hidden cost of giving an agent everything
Every tool you add gets paid for twice: in tokens and in hesitation. Fifty available tools do not make an agent more capable, they make it more indecisive.
Context engineering
Context engineering: the time a photo took an agent down
A client sent five pictures of their shop over WhatsApp and the agent stopped responding. The cause was not the model: it was what we let into the context window.
Let's talk about
your project.
Tell us what you need and we'll tell you how we'd approach it. In minutes, not weeks.
No strings attached. The first chat is free and we reply right away.