Field note 02 / 10
The Night the Agents Ate the Spec
Before Rosetta Bitcoin had a proof machine, a smaller system showed the dangerous promise: agents could move like a swarm if the spec, tests, and human intent stayed in the same loop.
The inciting incident did not happen inside Bitcoin.
That is the honest part.
It happened in a different codebase, on an ordinary night, with the kind of software people build every day and still somehow never fully specify. I had a large system, a structured markdown spec, a model acting like a compiler, and a test oracle that did not care how impressed I was with myself.
The experiment was not "can a model write code?"
That question was already boring. The better question was whether an agent swarm could operate around a precise artifact: one model generating, another checking, tests rejecting the fantasy, and the spec tightening until the output stopped being a vibe and started being software.
The setup was almost embarrassingly plain.
Write the behavior down. Let the agent generate. Run the tests. Read the failure. Tighten the spec. Run it again.
At some point the loop stopped feeling like prompting and started feeling like machining. The model was not the miracle. The spec was the jig. The tests were the measuring tool. The agent was fast, but speed only mattered because the work had a shape.
That night changed the frame.
If a careful specification could move that much software, what else had we been misjudging? How much of "engineering difficulty" was really ambiguity being passed around as labor? How many expensive problems were expensive because the artifact that should carry intent had never been made?
That question is dangerous. It can make a person overreach.
I overreached.
The service-manual idea for Bitcoin stopped looking impossible. Not easy. Not safe. Not something one person should trust alone. But no longer impossible in the old way.
The machine had shown me a door: if the spec is the artifact, and the artifact can be verified, then agents can become collaborators in a much longer kind of work.
But Bitcoin is not an ordinary app. It is not a private system with one owner and a sprint board. It is a living network protected by people, incentives, history, adversaries, and review culture.
The lesson from that night was real.
It was also incomplete.
The agents could eat the spec. The harder question was whether the spec could survive the agents.
That question followed me out of the smaller system and into Bitcoin. Speed was not the point. Speed without memory is just a faster way to lose the plot.