The Craft Before It Was Automated
The skill being built right now — how to coordinate distributed agents, how to structure information so any model can navigate it, how to design adversarial review into a research workflow — may not be rare for long.
The techniques that feel hard today will be automated and abstracted within a few years. Entry structures, belief registries, checkpoint patterns, multi-agent coordination, role-separated agents with adversarial framing — all of it will become platform features. Configuration instead of craft. The platforms will catch up.
The window is short.
What Gets Preserved
The people who built it by hand, who hit the failure modes directly and worked out the solutions from first principles, will have something the platform users won’t.
Not the techniques themselves — those will be commoditized. The intuition about why the techniques work. What breaks without them. What the failure looks like before you know what to look for.
You can’t get that intuition from a platform. You get it from watching five of six verification tests turn out to be tautological. From finding that every role definition file was stale after months of operation. From seeing a pipeline declare SATISFIED on code it had just acknowledged was buggy. From losing working state at a context boundary and having to reconstruct everything from scratch.
The failure mode experienced directly produces something that documentation of the failure mode cannot fully convey.
What the Documentation Is Actually Doing
This blog series isn’t just content. It’s a record of building the craft before it was automated.
Not “here’s how to use entry and beliefs” — but “here’s what we learned by building them, here’s the failure mode that forced each tool into existence, here’s why the sleeping dog wakes up at the wrong time.” The mechanism, not the conclusion. The experience, not the summary.
That record will be more valuable than the tools themselves once the tools are table stakes. Everyone will have the tools. Not everyone will understand why they work.
The Timing
This is the early period of a field being built in real time. The people working in it now are not just users of a technology — they are accumulating the foundational intuitions that the field will be built on.
The analogy: the Python developers of 2005 who understood why dynamic typing worked and where it failed, who internalized the cost model of the interpreter, who built intuition about when to reach for C extensions — they had something that the Python developers of 2015 who learned from tutorials didn’t. Not better skills. Deeper foundations.
The distributed agent coordination craft being built right now is in the 2005 moment. The platforms are coming. The window to build foundations rather than use abstractions is open now and will not stay open.
Post 7 in a series on the AI economic shift. Previously: The Power Gap Will Close. Series start: Software Is the Modern Rai Stone.