Philosophy
I’m passionate about agentic engineering. I love using agents to get shit done, and I want to keep getting better at it. Here’s how I’ve been formulating that:
A 10-out-of-10 model in a 2-out-of-10 harness gets you a 2, not a 6. You can’t average your way out: whichever term you neglect discounts everything the other two put in.
That’s the whole philosophy in one line. Everything else is looking at the terms one at a time, because each one asks something different of you.
Model: given
This is the term you get for free, and it’s the best deal in the industry. The model improves on its own, the releases keep coming faster, and every few months the exact same harness and skills get a stronger engine dropped into them. You don’t have to earn any of that. The only thing this term asks of you is to know the models well enough to pick the right one for the job.
Harness: yours to customize
The harness is everything between the model and the work: tools, context, memory, skills, permissions, verification. You don’t have to build one; off the shelf, Claude Code or Codex is already good. But stock is the most convenient setup, not the most effective one. Every hour you put into customizing compounds across model releases instead of resetting with each one.
You: the operator
The last term is you. Steering agents is a skill, and it responds to practice: try different models, try new techniques, read how these systems actually work, stay current with the tooling. The people getting remarkable results aren’t smarter. They’ve practiced more.