What do we still need source code for?

It is an odd question to ask after spending forty years writing it, but it is the one that keeps pulling at my sleeve. Let me work backwards to explain why.

The first computers were one-trick ponies. Their behavior was baked into their wiring — change the task, change the machine. Useful, expensive, inflexible. Then the elegant idea emerged that part of the data a machine processed could also control how it processed the rest, and the stored-program computer was born. Hardware became a stage; software became the play.

And with software came developers — a new profession whose first and hardest job was, and still is, to understand a process well enough that they could, in principle, perform it themselves. The encoding into source code was always the second step. We did it because humans, however well they understand a process, cannot match a machine for speed or reliability — and have an inconvenient need for sleep.

For decades that was the deal. A business owner understood a process; a developer understood the business owner; the source code was the byproduct of that understanding, painstakingly translated through several meetings, languages, frameworks, and rather more meetings on the way to silicon.

That deal has changed. The entity we now describe processes to is already the machine. The author and the audience have merged. So the question writes itself: if the agent already understands what we want, why do we still ask it to produce thousands of lines of source code that we, in turn, will mostly never read?

The short-term answers are perfectly good. Executing compiled code is cheaper and faster than burning tokens. The entire existing body of software — every library, every API, every running system — is encoded in source. That body of work is not going anywhere quickly — not in a year, not in a decade, probably not in two.

But mid-term, I think the answer changes. Our industry has a stubborn habit of making things cheaper and faster, fast. The obstacles ahead are real, but they are the kind of constraints we have spent decades learning to engineer around. Once the economics flip, the cleanest representation of an application is no longer a tree of source files written for one runtime — it is a single document, in prose, describing the logic and behavior the application is supposed to exhibit. Read directly. Understood directly. Acted on directly.

That is what I mean by speccing is the new coding. And it is the reason I have just published SpecPack — three small reference standards meant as an experimental foundation for that future. None of them are rocket science. I think of them as a bit of housekeeping for a fresh start.

MiniMark takes Markdown and removes its optionality. Humans thrive on optionality — it is how we express our individuality and our taste. Machines do not need it, and tend to find it actively confusing. MiniMark keeps the syntax humans already know and strips the redundant ways of saying the same thing.

riVer is a versioning scheme for textual content. It assumes a world in which sophisticated version-control systems like git are no longer required, or simply not present in the agentic environment. Once an application is a single document, a long-standing anti-pattern turns out to make sense again: putting the version number inside the document and hence into the agent's context. What that version should indicate, on the other hand, is a question we get to ask from scratch. Agents do not consult a semantic version to decide whether a change is breaking — they read the spec and find out. What they need is an integer to mark the iteration, a status to mark the lifecycle stage, and a timestamp to place the change in time. riVer gives them exactly that.

Specify is the most ambitious of the three: an attempt at American English coding standards — conventions for expressing programmatic logic and behavior in plain English, precisely enough for an agent to act on.

So — does this mean source code goes away?

No. It changes jobs: from the substance of applications to the seam between them. More on that next time.

In the meantime, the standards are live, the license is permissive, and the floor is open. Have a look around, and let me know what you think.