Most agent-velocity hype rests on one premise: that writing code was the slow part. .txt, the team behind the structured-generation library, takes a saw to that assumption. The point goes back to two foundational software-engineering texts—Fred Brooks’s The Mythical Man-Month (1975) and Gerald Weinberg’s The Psychology of Computer Programming (1971)—and .txt puts it like this:

Software is what’s left over after a group of humans finishes negotiating with each other about what the system should do. The code matters, but it is the residue of the harder work, not the work itself.

Code as residue. That inversion reorganizes the whole conversation. The tools and processes we’ve built around software for fifty years—IDEs, wireframes, mockups, code review, even pair programming—have been about lowering the cost of producing the residue. Once that cost approaches zero, what’s left to slow you down is the negotiation underneath. And that negotiation has not gotten any cheaper.

What that layer actually consists of, in practice:

What slows down a team where agents do the implementation is the production of specifications precise enough for an agent to pick up and run. Roadmap, written down. Acceptance criteria, written down. The “what we actually want” forced into precision, be it via a test suite, a ticket, or a written design.

The bottleneck moves from people writing code to people deciding what code should exist. .txt calls that work management, and I’d put it a little wider; it’s also product, design, and anyone whose job description includes the phrase “what we’re building.” A spec precise enough for an agent is a falsifiable description of the outcome, with the trade-offs already made.

.txt on what runs underneath the spec:

Context is the commodity an organization runs on. It is the shared understanding of what we are building, why it matters, what has been tried, who decided what, what is load-bearing and what is vestigial. Humans on a team accrete it by osmosis. By being in the room, by reading the same Slack channel, by debugging the same outage at two in the morning. Most of it is never written down. When a senior engineer reviews a PR and says “this’ll break the migration,” they are drawing on context that has no document. Agents cannot do osmosis.

“Agents cannot do osmosis” is the line. Specs are the formal surface; context is what’s underneath, and teams absorb it without writing it down. The post closes here:

The companies that win the next decade will not necessarily have the best models or the best agent infrastructure. It will be the companies whose fifty people, then two hundred, then two thousand, can stay aligned on a shrinking set of decisions while shipping more output per head. They will be the ones that already knew, before agents arrived, that their hardest problem was coherence. That is a culture and management problem. Always has been.