The bottleneck moved. Most people haven’t noticed.

For most of software’s history, the scarce thing was people who could write the code. That constraint is gone. AI writes most of the new code at the frontier now — three-quarters of it at Google, and every line still approved by an engineer.

So the work should be ten times faster and cheaper. It isn’t. Ninety-five percent of enterprise AI pilots show no measurable impact on the bottom line. Not because the models can’t code — they can — but because a demo that ships beautifully on a clean repo collapses the moment someone asks the only question that matters in a real organization: who approved this?

The bottleneck isn't writing code anymore. It's accountability.

A copilot makes one engineer faster. A task-agent takes a ticket and returns a pull request. Both are useful. Neither is accountable — there is no chain of approval, no tamper-evident record, no control over the blast radius. And so the enterprise, correctly, won’t let them near the thing that matters.

What’s missing isn’t a smarter model. It’s the organization the model works inside: identities, so every action has an owner. Gates, so nothing ships unproven. An audit trail, so the record is either intact or provably altered. A vault, so secrets never leak. A kill switch, so a human is always one click from stopping everything.

Anthropic sells the engineer's brain. We sell the organization the brain works inside.

We know it works because we watched it. One evening, a one-line product idea became a signed spec, a backlog, thirty-plus rounds of review between two agents with distinct GitHub identities, and a deployed product — in two hours and thirty-six minutes, with exactly three human decisions. Every step is on the hash-chained record; we couldn’t have faked it after the fact if we wanted to.

That’s the whole idea. Not autonomy for its own sake — autonomy someone can answer for. An engineering team that isn’t people, but is accountable like one.

Worst case is a rejected pull request.