There was a time when ships crossing the same ocean carried different charts of it. The thing that tripped them up was longitude. Latitude was relatively easy — you could read it off the height of the sun and the stars — but to measure how far east or west you’d traveled, you first had to settle one question: where do we put zero? And the answer differed from one nation to the next. A British ship took London as its prime meridian, a French ship took Paris, and some ship or other took the harbor it had sailed from. So the same island landed in a different spot on every chart. No one’s map was wrong, and yet none of the maps agreed.
It wasn’t until 1884 that the world settled on a single line — the one running through the Royal Observatory at Greenwich — as the prime meridian. Only then did every chart start speaking the same language. It was the moment a scatter of conflicting references collapsed into one reference that everyone agreed to point at.
People who work with software meet this problem every day. When the same fact is written down in several places, over time the copies quietly drift apart from one another. So there’s a design principle that says: decide clearly that the real reference for this information lives here, in one place. The single source of truth — SSOT for short. It used to be treated as a nice-to-have fundamental, but once AI agents moved to the center of the work, it abruptly became something that decides whether the work succeeds or fails.
An old dream of a single truth
A single source of truth isn’t some grand piece of technology; it names a state. For any given fact, there should be exactly one place you can genuinely trust as its origin. Everything else is copied from there, or generated from it automatically.
A bank passbook makes it easy to picture. The passbook doesn’t just store the result — “current balance: one million.” It stacks up the transactions in order: deposit 500,000, withdraw 200,000, deposit 700,000. The balance is just a number computed by adding those entries up. So the real truth lives in the transaction history, and the balance on the screen is a shadow derived from it. Even if the screen breaks, as long as the history is intact, the balance can always be reconstructed.
This arrangement falls apart the moment you copy the same fact by hand into several places. Fix one and forget the other, and the two quietly drift apart. Nobody lied, and yet the system as a whole can no longer tell you what’s true — which is to say, you’re back to the age of sail and its conflicting charts.
So a well-built system handles truth in two distinct kinds. The place where a piece of data is first born and where its owner lives is the system of record. Customer data is born in the CRM, accounting in the accounting system. And separately there’s a fast copy that mirrors the record accurately and serves it up — the system of reference. One record, many references. As long as the promise holds — that no matter how many references there are, the truth is the single record — there’s never any doubt about what to believe.
A person stops and asks; an agent trusts and acts
Why has this old principle become urgent again? Because the work is steadily passing out of human hands and into agents’. By “agent” I mean an AI that takes several steps on its own in a person’s place. And an agent takes facts in a completely different way than a person does. That very difference is what makes the single source of truth urgent.
A person hesitates when something’s murky. If a number in a document doesn’t match what they heard yesterday, they think hm, that’s odd and check it once more. Even when information is out of sync, the net of human common sense catches the big mistakes. An agent is different. It finds and reads the document, reasons from its contents, and moves straight on to the next action. Which means it takes the information it’s given as true and acts on it at once.
So when the truth wobbles, that wobble passes directly through into a wrong action. Think of the agent as a new hire who does exactly what they’re told. It works fast and meticulously — but hand it a faulty manual and it will execute the wrong thing flawlessly. The spot where a person would have paused and tilted their head, the agent sails right past, deploying, placing orders, deleting without a flicker of doubt.
There’s one more property that makes the situation thornier still. The same question can draw a slightly different answer from an agent each time. This is called non-determinism. When behavior is already this shaky and the information shakes too, even tracing back what went wrong becomes impossible. So at the very least, the information has to be nailed to solid ground.
Turn it around, though, and that’s exactly why a solid single source of truth becomes firm footing for an agent. It’s like handing that same do-exactly-as-told new hire one manual that never wavers: they work faster and more consistently than anyone. That manual gives an agent three things. The first is trust: with a single origin, the agent decides I just trust this and moves without hesitation. The second is grounding: fill in the blanks with no basis and you get hallucination, but answer from the given source and it holds up — what’s usually called grounding. The third is simplicity: when facts are scattered, you have to dig through them all and resolve the conflicts, and an agent can’t paper those over with common sense the way a person can.
The answer was already there in 1991
What’s striking is that this isn’t a new problem. The question — when several autonomous actors work while sharing the same facts, how do you keep the truth consistent and contradiction-free? — is in fact something AI research has wrestled with for more than thirty years.
In 1991, the AI researchers Michael Huhns and David Bridgeland published work on consistency among multiple agents sharing the same beliefs. The core was holding two kinds of consistency at once: each agent has to be coherent within itself (local consistency), and the facts everyone shares mustn’t contradict one another (global consistency). The name itself is truth maintenance. It’s essentially the multi-agent prototype of what we now call a single source of truth.
The roots run deeper. In 1985, three logicians laid out the theory of belief revision: a set of rules for taking in new information while giving up as little as possible of what you already believed, so the whole body of knowledge stays a single, contradiction-free state.
The approach being pulled back off the shelf to make multiple AI agents collaborate sits on this same lineage. Instead of letting the agents whisper to one another, you have them write only to a single shared board and read only from that board. It’s called a blackboard architecture, and it’s actually an old design out of 1970s classical AI. Making everyone look at the same single place shuts off, at the source, the problem of each actor harboring different facts and pulling out of line. It’s exactly the same idea as fixing the Greenwich meridian.
So the single source of truth isn’t a buzzword the agent era just coined. It’s closer to an old question, long ago thrown down, finally coming off the lab bench and into the work of the real world.
And so Git became the standard
An abstract principle has to be built into a tool in the end. And the tool best suited for the job right now is Git. It was originally made for programmers to manage code — and its properties happened to line up with exactly the conditions an agent wants from truth.
Three of Git’s properties become, directly, the requirements of truth. First, once a change is recorded, who changed what, when, and why is preserved forever (immutability). Second, to change anything, it has to pass someone’s review and approval (verification). Third, the state at a particular moment can be pointed to again, precisely, with a single tag (fixity). For an agent to be able to say I trusted this version and worked from it, it needs exactly these three things.
By analogy, Git resembles the passbook. The commit history is the transaction record; the current state of the files is the balance. That structure — the real truth lives in the change history, and the present is derived from it — maps exactly onto the ideal of a single source of truth. Git was always a good candidate to be the reference for truth; the agent era is what finally drew out that use.
And yet the single truth is a myth
Get this far and the natural conclusion seems to be: so just gather everything into one place. But this is precisely where a serious objection arrives. An absolute single truth is, in fact, closer to a myth.
Zhamak Dehghani, who created the design movement called data mesh, puts it provocatively: multiple sources of truth aren’t a bug, they’re a feature. Hauling everything to the center only grants the illusion of control — press one side of the balloon and the air bulges out the other side. Different departments legitimately see the same data differently. Even a single word like “revenue” gets counted differently by finance, marketing, and sales, depending on whether it’s before or after refunds, with tax or without. No one is wrong; their purposes differ. Force them all into one definition and, for most people, the number no longer fits their work and becomes useless — so everyone ends up building their own spreadsheet on the side to route around it.
Which yields one insight. Gathering data into one place is, if anything, the easy part. The genuinely hard part is getting everyone to agree on what that data means. Truth lives not in the data but in the definitions. A single store is a necessary condition for a single truth, never a sufficient one.
The real cause of failure, too, is usually not technical but political. Only I know this. My department’s number is the company’s official number. The turf war. A single-source-of-truth project tends to break not because the data can’t be gathered, but because people resist a demand to give up control.
Where the tools diverge
So the real-world solution doesn’t go to “everything into one,” but somewhere more subtle. Keep the record cleanly singular, but manage well the different representations and agreements that context demands. And this is exactly where the flood of recent tools diverges.
On one side are tools that lay a layer on top of where Git is weak. Git is unfriendly to people, so they put a pretty editor like Notion over it; it can’t handle tabular data, so they add version control built for data; agents can’t read meaning and relationships from it fast, so they weave the knowledge into a graph and feed it that way. On the other side is a current that copies what Git does well — immutability, verification, traceability — out beyond code, onto data and documents and AI assets. Either way, the starting point is the same: how do you build the solid floor an agent can trust and act on?
There’s an interesting paradox here too. The agent demands a single source of truth more urgently than ever — and at the same time lowers, on its own, the cost of maintaining one. Work that once required a person to sync documents by hand, an agent now does automatically, generating the derived copies from the record and keeping them aligned. The single source of truth that always rotted because it took so much hand-labor can, at last, be kept alive.
A standard that doesn’t move, again
The ships of the age of sail each held a correct chart and still disagreed. The problem wasn’t accuracy; it was the reference point. Only after a scatter of truths was gathered onto a single meridian could everyone finally sail the same sea.
The age of agents sets that old assignment back on the desk. Only, the answer isn’t the simple order “gather everything into one.” As the 1991 work already showed, the heart of it isn’t singularity in itself, but keeping the shared truth contradiction-free. Keep one trustworthy record; allow derivations as context demands; and watch, ceaselessly, that they don’t drift apart. The real goal is not the fantasy of one absolute truth, but the ability to manage divergence.
Which leaves one question. Who pays for the work of building and maintaining this standard that doesn’t move — and how much? The story of that cost continues in the next piece.
— tomte