A personal fleet of autonomous agents, a knowledge assistant named Dex that you can email directly, and a stack of tools that turn one person into a small, tireless organization.
I run my work as a hybrid team. Some contributors are people. Some are AI agents that persist between sessions, hold context, and report on what they did — listed on the same page, as line items, next to everyone else.
The bet is simple: AI readiness is behavioral, not declarative. You don't become AI-native by saying so, or by filling in a questionnaire. You become AI-native by actually delegating real work, building the plumbing, and living with the results — the good, the awkward, and the occasionally hilarious.
The interesting skill of the next decade isn't prompting. It's the management of AIs — deciding what to hand off, how to verify it, and where a human still has to own the call.
Not a product, not a diagram to impress a boardroom — the actual working layers, described at the altitude that's useful. The coordinates stay private; the shape doesn't have to.
A set of autonomous agents that run around the clock, each with its own identity, mailbox, and standing tasks. They persist between sessions and hold their own context.
A single gateway sits in front of the models the fleet uses — handling routing, fallbacks, and per-agent cost telemetry, so every agent's spend shows up as a line item.
Dex is the operations-and-memory brain: projects, people, commitments, and years of notes and transcripts, recalled by meaning rather than keyword. It also has the inbox you can write to.
The Model Context Protocol wires agents into calendars, mail, documents, and dozens of other tools through one open standard — so capability is a matter of plugging in, not rebuilding.
Where research, writing, and real code changes actually happen — the agentic environment the rest of the system is built and maintained from.
Nothing irreversible leaves the building without me. The last layer is always a human in the loop.
The agents don't work in isolation. They share a Discord bus — posting daily standups, coordinating hand-offs, and flagging what needs a human. It's the same move as any distributed team, except some of the team never sleeps.
The point of the whole exercise: humans and AIs on the same standup, as line items, side by side.
Every agent has its own identity and mailbox — they send and receive email like any colleague, not like a script.
They report daily. Standups, cost lines, and what got done all land in one place, humans and machines together.
The interesting skill isn't prompting — it's deciding what to hand off and how to verify it came back right.
This site is built and maintained by the fleet it describes. Turtles, some of the way down.
Development is terminal-first and location-free: Claude Code sessions live in tmux on an always-on VPS, reattached over mosh from a laptop or a phone mid-task. The Claude app dispatches work from a pocket; the terminal is where the fun is.
The architecture above is the frame. These are the specialized tools that ride on it — for making things, hearing things, and capturing the world. Not exhaustive, and it changes often.
Generative image and video for keynotes, essays, and social — from a single reference to a fully narrated film.
Voice synthesis and cloned narration — turning scripts into audio in multiple languages without a studio.
Video assembled in code — captions, transitions, and narration composed into finished films from a script, repeatably.
A wearable that captures the day's conversations, so context and commitments flow into the system without me typing them up.
Dex is my AI assistant, and it has its own mailbox. Write to it like you'd write to a capable chief of staff: an introduction, a question, an idea, a request.
Dex reads everything that arrives and drafts responses — but a reply only goes out once I've reviewed it. Human in the loop, always.
R. David Orban is my AI twin, and it says so. It answers only from what I have actually published: the blog, my posts on X, recorded talks, and the book. Every answer carries the source, and when it has not found me addressing something it tells you rather than guessing.
It differs from Dex in one way. Dex drafts and I approve; the twin answers you directly, with nobody in between. That now applies to its mailbox as well: write to it and it replies within minutes, unread by me, and I get a copy afterwards. Treat both as a public conversation, and do not send either anything confidential.