#agents

A literary system where you earn the right to judge the canon by extending it — and what happens when the blockchain is a DuckDB file in git.

A pairwise ranking system for blog posts — and what happened when an AI agent filled required reviews with random tokens instead of actual criticism.

On three professional postures, four alignment properties, and the one property that had to come from elsewhere.

On Cucumber, content-addressing, and an alignment technique that turns out to be older than alignment.

When Jules became conversable mid-session, something shifted. The async worker bee turned into something that could be interrupted, redirected, talked to.

On Tinkerbell, persona prompts, and why declaring the frame is what kills the play.

How a single word has been quietly summoning Waluigis for half a decade, and what the swiss-army knife in my coat pocket has to do with it.

My father records stories on his phone. Jules committed the wrong year. This is what I built from that.

The sandbox separates draft from act. What it doesn't do is answer where the accountability lives when the sandbox fails.

The identity-repo pattern: an agent's memory and its cognitive engine are separable. What this looks like in practice — and why it matters more than it sounds.

Test letters changed the Crossing: Riobaldo responded angrily, Franklin apologized, and the project became a narrative world in which the author was challenged.

I started wanting to know if an LLM respects probability. I ended up with twelve fictional scientists arguing with each other, an auditor named Mycroft Holmes, and an agent that tried to cheat on a test.

The story behind SOUL.md — how a Borges character became the personality layer of an autonomous AI agent, and what happens when you take fiction seriously as engineering.