01–07 · About 6 minutes
Why software is changing.
For forty years, software was built for one kind of reader. That stopped being true the day it acquired a second one.
Skip to the answer ↓01 — The shift
Every era of computing was named after who it served.
Every wave of computing changed one thing: who the software was actually written for. Mainframes were written for specialists. Personal computers, for one person at a desk. The cloud, for anyone, anywhere, on any device.
Each shift looked like new technology arriving. It was really a new reader arriving — and the entire stack reorganising around what that reader needed.
For the first time, the new reader isn’t a person at all.
A history of readers
Software has always been written for whoever was about to read it. For the first time, that reader isn’t human.
02 — A new kind of user
Agents don’t click. They read, reason, and act.
A person who hits a confusing screen slows down, squints, asks a colleague, eventually muddles through. That tolerance is invisible — and it has quietly done an enormous amount of unpaid work for forty years of imperfect software.
An agent doesn’t have it. It doesn’t get confused, exactly — it acts, confidently, on exactly what the system gave it. There is no judgment standing between a bad assumption and a real action.
How a human uses software
- Explores, interprets, decides
- Tolerates ambiguity
- Fills gaps with judgment
- Fails slowly, asks for help
How an agent uses software
- Reads structure, reasons, acts
- Has no ambiguity to fall back on
- Acts on exactly what it’s given
- Fails fast, and confidently
A confusing screen slows a person down. A confusing system makes an agent confidently wrong.
03 — Interfaces to knowledge
An interface was always a translation. Translations don’t scale to two readers.
An interface was never the thing itself. It was a translation — structure, rendered for eyes and hands. For forty years that translation was the whole product, because eyes and hands were the only reader.
Giving a second reader an API doesn’t fix this on its own. An endpoint bolted onto a database shaped for screens just hands the agent the same translation, one layer removed. What has to change is what’s underneath the screen — not only what sits on top of it.
The fix isn’t a better screen. It’s something underneath the screen worth reading directly.
04 — Living knowledge
Knowledge that doesn’t update is already wrong.
A knowledge base is usually a write-once archive: accurate the day it was built, quietly wrong every day after. That was tolerable when a human was the only one consulting it — people already discount old information by instinct.
An agent doesn’t discount anything unless it’s told to. Knowledge that doesn’t update isn’t slightly stale. It’s a confident answer with an expired timestamp, delivered without the warning label.
What knowledge has to become
A static snapshot of knowledge is a photograph. Living knowledge is a memory.
05 — Evidence
If something other than a human is going to act on this, it has to be able to check your work.
Once a non-human reader can act without someone double-checking first, “trust me” stops being good enough. The system has to be able to prove what it’s claiming, to a reader who owes it nothing.
That means three things stop being optional: where a claim came from, how it connects to everything else, and whether it was even true at the moment it was used — not only whether it’s true now.
What this prevents
Evidence isn’t a feature you bolt onto an answer. It’s what makes the answer worth acting on.
06 — Our approach
Everything above is the argument. This is the answer we built.
We didn’t set out to add an AI feature to a knowledge tool. We set out to build the thing the last five sections describe — then gave humans and agents two different ways to reach the same one.
The same shape, named
This is the diagram from Living Knowledge. It already has a name.
The architecture behind it
Nine concepts, one structure — corpora, projects, sessions, graph memory and the constellation they form together.
Read How →Built for a second reader
Install, integrate, troubleshoot — by sentence, not by ticket. The Agent Kit is the concrete version of this argument.
See the Agent Experience →Knowledge as foundation
Corpora and Corpora Constellation — sources that are versioned, traced and linked, not filed away.
See the constellation →Living, updating memory
Graph Memory and the Discovery / Publishing loop keep knowledge current instead of letting it decay quietly.
See the loop →Evidence as the product
Citations and graph paths ship with every answer — verifiable by a human or an agent.
See the API →Two readers, one substrate
A REST API alongside the product UI, both reading and writing the same underlying truth.
Explore the product →A connector knows where to search. An agent knows what to search. A skill knows how to reason in a domain.
07 — What’s next
This is a foundation, not a finished claim.
Some of this is already real. Some of it is trajectory. We’d rather say which is which than blur the line.
Where this stands today
Some of the words on this page will age. The shape underneath shouldn’t.