Blog
Notes from the practice: what the research papers actually say, what we shipped to test them, and what we'd build with another week.
Practice
The agentic-engineering practice.
The methodology I run, published in public. Definitions and the working tools.
What "agentic engineering" actually means.
Agentic engineering is the practice of becoming proficient with AI coding agents and shipping systems that can be externally validated. You're learning what the agents do well, what they don't, and how to design the harness so every change is testable, reviewable, and reproducible — not just believable. The team that ships safely in 2026 is the team that ships what they can prove, not what the agent told them.
It is not an "AI agency," a "prompt-engineering shop," or any other rebrand of "we use ChatGPT." The agents are tools. The engineer uses them. The shipped code carries the engineer's name.
The three practices
-
01
Specify first.
No agent run without a written contract. The spec names the inputs, the expected outputs, the tests that will judge the result, and the failure modes that mean "stop and reframe." Vague prompts get vague code.
-
02
Verify, don't trust.
Every change goes through automated checks, a human reviewer, or both. The harness produces the receipt; the engineer reads it. "It compiled" is not a verification — it's the input to one. Receipts, not vibes.
-
03
Encode expertise.
Equip every agent with high-quality skills, persistent memory, the right tools, and zero-token architecture for routine work — so the engineer reviews judgment, not boilerplate. The harness carries the playbook; the engineer carries the on-call.
What I publish
Public artifacts.
A practice unpublished is just a habit.
Research skills
Open-source skills for AI agents — the working tools the studio uses to research, plan, and ship.
Tell me what your agents shipped last week.
Or the repo your team is afraid to touch. I read every message myself.