The Urban Design Of My AI Agent Architecture

With AI agents rapidly gaining new capabilities, I created a system that I find very useful. You can find its description on ai.davidorban.com.

A set of persistent agents run around the clock, each with its own identity, its own mailbox and its own standing tasks. A single gateway sits in front of the models they use, so routing, fallbacks and cost telemetry are one problem rather than one per agent. Dex is the operations and memory layer, holding projects, people, commitments and years of notes and transcripts, recalled by meaning rather than by keyword. It is based on a fork of the Dex open source project. The Model Context Protocol wires all of it into calendars, mail, documents and several dozen other tools through one open standard, which turns new capability into a matter of plugging in.

The animation of the agents and their relationships is a useful metaphor of urban design. How they group by function, how information flows between them. As their number increases by orders of magnitude, which I expect over the course of the next few months, new relationships, and emerging structures will become visible, and extend the metaphor.

Two of the agents face the public, and they are governed in opposite ways.

R. David Orban is my digital twin. It answers from a knowledge base of my published work: over 500 blog posts, more than 300 recorded talks, twelve thousand posts on X, and the text of my book Something New. Twelve hundred files, about 1.7 million words. When the corpus does not contain an answer, the twin says it has not found me addressing the subject and stops there. Those refusals are collected, grouped by theme once a week, and become a queue of things I have never written about, but my community is interested in, and lets me decide what to cover next, enriching my thinking.

Until last week the twin only held conversations on the web, which you can still do, either from the home page of my website or from a dedicated page on twin.davidorban.com. But it now has a mailbox as well, at [email protected]. You can write to it, and it will reply within minutes. I get a copy of every exchange, and that is the whole of my involvement.

Dex has a mailbox too, at [email protected], and it works the other way around. Dex reads what arrives and drafts a response, and the response waits until I have read it. Nothing leaves that account, and no action outside it, without me saying so in a separate step.

The asymmetry is deliberate and it is not about trust in the two models, which are much the same model. It follows from what each agent can reach. The twin can only assert what it can retrieve from a fixed body of published text, so the worst outcome of a hostile email is a wrong or unhelpful answer sent to the person who wrote it. Dex sits on the calendar, the documents, the files, so the worst outcome of a hostile email is an action taken in my name. The first risk is bounded by construction. The second is bounded only by a person looking at it.

That is also the argument of a paper I published this summer, which makes it a useful test of whether the argument survives contact with my own infrastructure. Civilizational Metamaterials proposes that governance be treated as an engineering discipline with quantities in it rather than as a normative one, and it offers a constitutive law for coordination:

R_eff = β · (1 − ρ) · (1 − τ) · (1 − γρτ)

β is the branching factor of decisions, ρ is provenance fidelity, τ is the rate at which outputs are verified, and γ captures how far provenance and verification catch the same failures. The model predicts a sharp transition between a self-healing regime and a self-destabilizing one. The practical content is that provenance and verification are substitutes across a range: raise one and you can lower the other without crossing the threshold.

The two mailboxes are that substitution, built. The twin is a context-binding provenance mechanism in the paper’s taxonomy, and its ρ is high because it is structurally incapable of speaking without a source, which is what makes running it at τ near zero defensible. Dex has the broad interface, so its ρ against arbitrary inbound text is low and its τ has to be near one. Same operator, same week, two settings, and the setting chosen from what each agent can reach rather than from how much I happen to trust it.

The general form of the rule is that an agent’s autonomy should be set by the provenance of what it can say and the reach of what it can do, and that both of those are properties you design.

The system is described at ai.davidorban.com.