AI-native engineering, where agents are colleagues — not chat windows.
I build engineering organizations where AI agents do real work — reviewing code, writing docs, operating desktops, running deployments — each with its own identity, policy and audit trail. The infrastructure underneath is sovereign by design: the keys, the data and the record belong to the people running it. Before that, a couple of decades of security work for people who learned the hard way what dependency costs. This site is a working notebook and a home for the systems I ship.
Current areas of work
Designing the environments agents do their best work in: semantic interfaces, deterministic context, closed feedback loops.
Autonomous agents as accountable colleagues — real cryptographic identities, policies, and audit trails, and no raw keys.
What agents need to remember — reference truth, episodic recall, institutional doctrine — as separate, provenance-tracked services.
Deployments, review and shared state as signed events on infrastructure you own — platforms optional.
- 01Bring your own protocol
Meta's Muse Glimmer runs a frontier-ish agent on a single consumer GPU. So I built a harness to ask the only question that matters to me: can it do nostr — and can a $1000 box of old cards do real agentic coding? A field report on open weights, an unfamiliar protocol, and the orchestration trick that makes cheap models punch up.
- 02Before the Swarm
Part one of a series on Nostrc, vibe-coding, and finding out where the machines break
- 03Running the frontier at home
What it actually takes — in gigabytes and dollars — to self-host open-weight models in mid-2026, from Kimi and GLM down to the 20B models eating the low end.
- 04Agents that keep their promises
Give an agent a memory and it gets smarter. Give it commitments and it gets character.
A Nostr-native deployment and runtime control plane. It tracks your builds, deploys your containers, and tells you when something goes wrong.
A NIP-46 Nostr bunker for agent fleets. Every agent gets an identity; no agent ever sees a private key.
Automated code review for NIP-34 Nostr repositories. Local LLMs read your patches; nothing leaves your infrastructure.
- Autonomy is earned
An authority model for agent fleets: promotion by boring reliability, demotion at the speed of doubt, and a break-glass lever for 2am. Draft.
- The sovereign engineer
On owning your tools, your keys, your data, and the systems that shape your work. Draft — being rewritten from the actual practice.
- Agent experience
The environments agents work in matter more than the models working in them. Notes toward a discipline. Draft.
Available for talks on building AI-native engineering organizations, agent experience, and running fleets of AI agents you can actually audit.
Speaking →Most of what I build ships as an open repository — and much of it was built alongside the same kinds of agents it exists to support.
GitHub →