Look for the bottleneck in an organisation adopting agents, and the first one you find is rarely the model. It's what people believe the model can't do, what they think they must still do themselves, and what the organisation still rewards.
Our dark factory started in July 2025. So did others, with access to the same frontier models. Some of them stalled. Ours is still growing.
If the model were the constraint, everyone with the same model would hit the same ceiling. They don't. The difference is in what the organisation did around the model: the structure, the loops, and the beliefs that decided which of those got built.
“We'll wait for the next model” is a decision to keep the current constraint, and to lose the slow way.
Each of these sounds like caution. Each one, held as a fixed rule, caps throughput at human speed, and each one has an engineering answer that the rest of this book describes.
None of them is foolish. They were true of the process we had. The work is to notice when a belief about the old process is being applied to the new one.
One bad agent PR becomes the story that proves the whole approach doesn't work. The human PR that caused last quarter's outage is just a bug.
Hold both to the same standard: defects per change, and how fast the system learns from them. That's the comparison All software fails somewhere sets up, and it's the only one that tells you whether you're getting better.
If promotions still reward the heroic all-nighter, the clever hand-written module and the number of PRs reviewed, people will keep doing those things, however many agents they have.
Change what's measured and recognised, and the mindset follows. Leadership has to go first: a manager who says “try agents” and then asks why the PR wasn't hand-reviewed has told the team which rule is real.
| Stop counting | Start counting |
|---|---|
| Lines of code, PRs opened | Changes in users' hands, lead time |
| PRs personally reviewed | Defect classes made impossible |
| Heroic fixes | Escapes turned into checks |
| Being the only one who knows | Knowledge encoded as skills and rules |
| Hours worked | Constraints removed |
Orchestrating agents is a skill, and new skills follow a J-curve. The first weeks are slower than working the old way. That's where most people decide “it doesn't work”, and where a mindset of experiment instead of verdict matters most.
It takes hundreds of hours to get good, and thousands to get fluent. Give people the time and the permission, on purpose: protected time, safe codebases to try things on, and a team where sharing a failure is normal.
Illustrative shape, not measured data.
Same model, different organisations, different results. Name the beliefs that cap throughput, judge agents by the same rate you judge humans, reward the work that removes constraints, and protect people through the dip in the learning curve.