Writing
Notes from building software, debugging systems, and working out where AI actually holds up.
Featured
Case study
Chasing frontier-level coding on one laptop
KRYN is a local coding workspace for Apple Silicon aiming at frontier-adjacent quality on a 9B model. How it's built, how v1 shipped, and the benchmark rig that will say when it's there.
Case studies
Systems I built, how they failed, and what the evidence showed.
A voice agent that takes the call alone
Inside a production insurance intake agent: a realtime speech model, a Deepgram transcript beside it, the code that owns every consequential moment, and the caller as last check.
The model doesn't get to write history
A small local-first civilization game built on the rule I keep using for agents: models and interfaces may propose, but one deterministic layer decides what is true.
The destination is the only witness
An engine that gives an agent's irreversible calls an identity the model can't regenerate, a real state for not knowing, and a benchmark that could prove it unneeded.
DaVinci: coding agents are mostly plumbing
A multi-agent orchestrator that drives a task through research, planning, execution and integration tests, with human review only at the gates that matter.
Essays and field notes
Arguments and shorter observations, newest first.
Agents need a nervous system, not a bigger brain
The next gains in production AI agents come from durable execution and replayable state, not bigger models.
Human-in-the-loop is an architecture, not a checkbox
Real human-in-the-loop means escalation ladders, fail-closed gates, and a deliberate decision about when the human is the fallback and when the human is the product.
The model was speaking fine and lying in JSON
A production voice agent kept dropping calls mid-sentence, and the bug was in the tool-call arguments, not the voice.