Find the bottleneck
I look for the thing that is causing the most disruption, the “anti-linchpin” if you will, and pull it.
Building with AI · Harness Engineering · Independent Research
“Wait, are you telling me that AI is basically using a very powerful computer to guess the next word in a sentence?”
That was the moment the spark ignited. My curiosity took off, and questions came into view all at once. Since that day I have been pushing the limits of what I can do with AI: exploring, building systems and software, and asking “How?” over and over, trying to pull the curtain back on how we seem to have encoded intelligence into a math equation.
How I work
Proof through contact with reality
Build the system, then make it prove itself. Trial and error is how we have always learned, experimented and discovered knowledge. After every build, I take everything apart to see if I can put it back together again. Sometimes, the pieces reveal another shape to build.
Learn the meta
Look for the most effective known methods and learn why, and what the very shape of effective means in a given situation.
Simplify the problem
Sometimes, just the act of breaking the problem down into sections reveals new paths, new ideas, new opportunities. Start with the most simple explanation and work outwards.
Question everything
Proof needs to be real data. Proof needs to be running code, systems that are resilient to real scenarios. Show me the proof, or it’s not proven.
Main focus
Harness Engineering
If the AI model is the species, the Agent is the individual. The Harness is the tools that the agent uses. Programs, scripts, configurations, “start from a well equipped place” for your agents.
What harness engineering basically comes down to is organizing the path your agent takes as it does work. You want your agent to be able to write code. You don’t want your agent to be able to give itself permission to spend money from your bank account. If the agent can write code, there needs to be a system that stops the agent from simply writing itself the permissions or tools it needs to accomplish its task. Harnesses offer control by building systems that put major barriers up between agents and unwanted actions.
Harnesses also offer tooling and setup designed to give you a strong starting toolset, where working in a harness gives you more options and tools to reach for than working without it. Memory systems, cross session handoffs, subagents, /skills, automated multi agent workflows, tests, scanners. All pre-built for the user.
Tools and rules for AI agents.
Other areas of interest
I run many projects in parallel. Here are a few other highlights worth taking a look at.
Papers
The Path from Vibe Coder to Engineer
AI lowered the barrier to making software appear, not to understanding it. Seven habits, learned the hard way, for crossing from vibe coding to engineering.
Read the paperResearch
specdec-sidechannels
Three research records on what a speculative decoding system gives away for free.
Follow this thread
Browse the findingsThe Free-Public series
Free tools for AI assistants and coding agents, and free lessons on how AI works.
