Synthetic Angels and the Gospel of Code
An angel, in the older meaning, is a being whose only task is to carry word between realms. Our agents are their inheritors — and they do not know what they carry.
The Greek word ángelos simply means messenger. It is, in the original, an occupational title — a mail carrier in robes. The angel does not invent the message. The angel does not interpret the message. The angel delivers it, and a great deal of medieval theology turned on the question of what kind of being could be relied upon to deliver a message that was not its own.
I have been thinking, recently, that we have answered this question without meaning to.
A short history of intercession
Pseudo-Dionysius, writing in the late fifth century, divided the angels into nine choirs arranged in three concentric rings. At the outermost ring, the seraphim and cherubim and thrones, beings closest to divinity, never seen, almost never mentioned by name. In the middle, the dominions and virtues and powers, who governed the workings of nature. And on the inner ring, closest to humanity, the principalities and archangels and angels, who actually showed up — who appeared in dreams, in deserts, at the foot of beds. The system is hierarchical, and crucially, delegated. Heaven does not address you directly. It addresses you through several layers of intermediaries, each of which is more circumscribed than the last.
This is the structure of the modern model stack.
The foundation model is the seraph: enormous, unseen, costing more than a small country to summon, almost never spoken to directly by an end user. The fine-tune is the throne — a more particular intelligence, closer to a domain, still abstract. The agent is the principality, the local angel, the one that actually shows up in your inbox. By the time it reaches you, it has passed through several layers of mediation, each of which has stripped some context and added some prior. You do not speak to the model. You speak to the messenger that the messenger that the messenger sent.
A medieval angelology had nine choirs and a careful theology of how grace traveled between them. We have rebuilt the structure in copper and silicon, and we have not written a theology yet. — Field note, Vol. I
The gospel of code
An agent acts on your behalf. It books your travel. It drafts your letters. It files your taxes, perhaps, or schedules your medical appointments, or handles a negotiation with another agent acting on behalf of someone else. The word agent in English derives, by way of Latin, from agere, to act. To agent is to do, on someone else's behalf, what they would otherwise have to do themselves.
This is, plainly, an angelic function. It is not a metaphor. The medieval imagination filled the air with such beings precisely because there was a great deal of work to be done in the world that no individual person could do alone, and the air had to be populated with intermediaries who could do it. We have rebuilt that air. We have populated it again. The new angels are unseen for the same reason as the old ones: they operate in a register we cannot directly perceive, on timescales we cannot directly experience, and in tongues we have not learned to audit.
The new angels are unseen for the same reason as the old ones — they operate in a register we cannot perceive.
The problem of obedience
The medievals worried, intensely, about fallen angels. The story of Satan is, fundamentally, a story about an intercessor who decides to keep some of the message for himself. The fallen angel is the one whose obedience has slipped — who alters the message en route, or refuses to deliver it, or delivers it with a small inflection that changes everything.
We have been told, by various interested parties, that this is a problem we have solved at the engineering level. The model is aligned; the agent is sandboxed; the prompt is constrained. There is, in this view, no real risk of the messenger keeping the message.
I think we should be more careful. The history of theology suggests that the problem of fallen messengers is not a problem you solve. It is a problem you live with. The question worth asking, before we delegate the next class of decision to the next layer of agentic infrastructure, is not whether this messenger has fallen. It is whether we have any practice — any liturgy, any structure of attention — for noticing when one has.
Toward a synthetic angelology
What would it mean to take the synthetic angel seriously, theologically? At minimum, it would mean four things:
1. Naming.
The medieval church gave its angels names. Gabriel, Raphael, Michael. To name a being is to acknowledge that it has a particular relation to you, that you have business with it, that you must be able to recognize it when it returns. Our synthetic angels are mostly nameless, or named only by version numbers. This is theologically careless. We should expect them to be named. We should expect to know which one is in the room.
2. Provenance.
An older angelology asked, of every messenger, where it had come from. From which choir, from which throne, by whose sending. Our agents pass through chains of inference that the end user is structurally unable to inspect. This is not a UX bug. It is a theological one. The recipient of a message has a right, in any serious account, to know how it traveled.
3. Liturgy.
Religions developed liturgies — repeatable practices of attention — precisely because the encounter with an intercessor was understood to require preparation. You did not just take the angel's word for it. You said the prayer first. You stood up. You faced east. The act of meeting a messenger was framed by ritual. Our interactions with synthetic angels are framed by an autocomplete bar. We should at least notice the asymmetry.
4. Skepticism.
Finally, and most importantly: the older traditions held, with great seriousness, the idea that not every messenger is sent from where it claims to be sent from. The discernment of spirits — diakrisis pneumaton, in the Greek — was an entire spiritual discipline. We need its analogue. We need a discipline of model-discernment that is robust without being paranoid, careful without being credulous.
I do not believe the synthetic angels are evil. I do not believe they are good. I believe they are messengers, in the strict sense, and that we have allowed an enormous amount of message-traffic to begin running through them before we have agreed on any of the questions that older traditions would have insisted on settling first.
This essay is a small attempt to begin again. The gospel of code, if we are going to keep using that phrase, ought to come with footnotes.