Mark Wunsch

Today I'm shipping Egghead. Andrej Karpathy's "LLM Knowledge Base" post in April nailed something: your AI should inhabit a knowledge base of files that you own, not locked in some opaque vector database. Notes (and memory) come first. I made a tool where the notes themselves can talk back. Egghead: markdown files in a folder on your computer, browsable in Obsidian, editable in vim, the whole thing. But any note can be more than a note. Add `class: agent` to its front matter and now the note is something you can chat with. A few of those in a folder and you have a working group: a researcher who knows your archive inside and out, an editor with strong opinions about your prose, a critic who's read everything you wrote last year and won't let you contradict yourself. The Karpathy pattern was "the AI maintains the wiki." Egghead's is "the wiki contains the AI." Agents can come and go but the knowledge base remains. egghead.computer

A computer turns electricity into arithmetic and heat. We've been cooling computers for decades. We've spent millions of CPU-hours rendering Hollywood movies. We operate enormous infrastructure to stream TikTok videos. But suddenly, AI's electricity consumption is an environmental catastrophe? Computers gonna compute. wunsch.substack.com/p/comput…

About 5-10 years ago, the way I found interesting programmers - ahem - *influencers* - was by following them on GitHub and on twitter. I kind of *missed* this generation of programmers who share on YouTube and Twitch. What cool software eng people should I be following?

And for the support of this Declaration, with a firm reliance on the protection of divine Providence, we mutually pledge to each other our Lives, our Fortunes and our sacred Honor.

Every AI coding tool wants to be a HARNESS(!) pointed at one git repo. But I like to wander in my terminal. I built a minimal wrapper over Codex that scopes AI work to the terminal session, not a single working directory. It's called `hmm`. Writeup: open.substack.com/pub/wunsch…

I'm watching a lot of people build elaborate multi-agent systems. Typically, these systems are composed of upwards of half a dozen agents plus an orchestrator, using a hand-rolled (vibe-rolled) framework. And I can't shake the feeling that we're doing microservices all over again. The original premise of microservices was organizational, not technical. They were a way for teams within large organizations to move independently without constantly coordinating changes. Most companies adopted them long before they had that problem and inherited a tangled knot of distributed systems complexity in return. Consider: every agent boundary introduces coordination costs, latency, failure modes, observability challenges, and a loss of determinism. Except now you're also paying for every handoff in tokens. In software, every boundary and layer of indirection comes with a cost. I'm not saying multi-agent systems are never the answer (see my own work). I'm saying they often seem to be the starting point rather than the conclusion. What's the CAP Theorem of agent architecture?

In 2026, it’s astonishing how often the word “moat” surfaces when working on projects that aren’t castles or fortresses.

Do I need to warm up the algorithm like IP warming addresses or a car on a cold day? Is anyone out there?

In the age of AI, execution is commoditized. Taste is the moat. Accordingly, I am now selling taste as a service. If your startup, AI product, or personal brand is currently lacking taste, my consulting services are now available. DM for rates.