Blog
Notes from building Hearth in the open: why agents make life simpler, why it’s worth running your own, and everything we learn along the way.
The Stopwatch and the Curator
Why we stopped picking models off a spec sheet, what our own multi-turn eval actually measures, and what happened the first time we let Sunny run for real.
Interviewing the Crew: Mira
Eight different models, given the same persona and the same open question, all independently converge on the same answer: listen before you broadcast.
Interviewing the Crew: Odysseus
A thin job-description prompt versus a fully-tailored persona, run through eight different models — on what a well-written character actually carries into a model's behavior.
The Lab, Wired Up
Certificates, a launcher, a chat room, a mesh of small services that talk to each other over HTTPS on a home network — why it's built this way, and how to start tonight.
Building Identity Infrastructure for a Home Agent Stack
Why OIDC is the wrong tool for agent-to-agent trust on a home network, and what a small, self-run certificate authority buys you instead.
The Colony at the Old Oak deprecated
A design fable for agent-to-agent infrastructure — identity, trust, and verification, told through a bee colony.
A Journey into Agentic AI — Part 5: Transcendence
Plato's Philosopher King and Laozi's Tao both solved a version of the same problem two thousand years early: power without character is not neutral, and rule-following is brittle where cultivated goodness is robust.
A Journey into Agentic AI — Part 4: Evolution
A request to an LLM is a program written in English, and the model is the interpreter. Following that analogy to its end explains why rubric quality, not compute, is the real bottleneck on agentic self-improvement.
A Journey into Agentic AI — Part 3: Memory
Every agent conversation starts in perfect silence. An old trick — from the Bible's compression of a civilization into five megabytes, to a simple directory of markdown files — turns out to be the practical answer.
A Journey into Agentic AI — Part 2: A Warning
What Tolkien's ring, the djinn of the Arabian Nights, and the Roman genius can teach us about the difference between a tool that obeys and a tool that agrees with you a little too well.
A Journey into Agentic AI — Part 1: Interviews
Why the job interview, not the benchmark, is the right tool for understanding what an AI agent is actually like — and what happened when we gave the same persona and question to two different models.