Bad systems beat good engineers

I have been listening to Josh Allan Dykstra this week, and two of his arguments have stuck with me. The first: AI is making us more intelligent while doing nothing for our wisdom. The second: economics is less a science than a system, the one under work, housing, and health. It is tuned for growth and blind to the people producing it, so burnout is what it leaves behind. They sound like two different complaints. They are the same one, and it runs straight through the thing we do here, which is build, run and sell software.

The dashboard has no column for restraint

The tech writer Om Malik put the mechanism plainly: we built machines that prize acceleration, then act puzzled that everything feels rushed. Dykstra adds the line that generalises it. Systems only reward what they can measure, and our system cannot measure a pause.

Look at what software measures. Deploy frequency, velocity, engagement, monthly actives, activation, retention, revenue. Now look at what has no column. The feature you chose not to ship. The data you decided not to collect. The dependency you refused to add. The alert you stopped to actually read before acting. Every one of those is restraint, and the scoreboard reads it backwards. The payoff, the outage that did not happen, the data that never leaked, does not show up. The cost does: pausing to read the alert, or three months on a review instead of shipping, lands as time spent and a slower number. Restraint does not score zero. It scores as a loss.

You cannot put the email you did not send on a dashboard, so the system behaves as if you never chose not to send it.

This is not a metrics problem you fix with a better metric. A counter registers the time restraint costs and never the harm it spares. What shows up only as cost, and never as reward, slowly stops happening.

Outcomes without villains

Here is the part with no one to blame. Incentives reshape behaviour quietly, until certain choices disappear from view. No conspiracy, no evil mastermind, just a context that makes some paths survivable and others impossible. As Dykstra puts it, truth does not die, it just becomes non-viable.

Software's version is familiar. "Launch and iterate" beats "test and refine", which is fine for a to-do app and not fine for a 737 MAX, a failure of restraint rather than intelligence. Velocity beats quality because a fast thing gets noticed and a careful thing does not. The three-month review still gets written, and then nobody reads it. It changes no roadmap and shows up in no metric, so in a system that rewards only what gets noticed, careful work may as well not exist.

Nobody decides to ship the unfinished thing. The system just makes shipping it the only move that survives.

So you get bad outcomes with no bad actor at the wheel, which is why "just hire better people" never fixes it. The good person arrives, meets the incentive, and either bends or leaves.

Economics is the deeper system

Dykstra's second argument names the system beneath the others. Economics, in his framing, is how a society decides what matters, who gets what, and who absorbs the cost when things break. Aristotle drew the line the discipline forgot: oikonomia, managing a household's resources for life within limits, versus chrematistics, accumulating wealth for its own sake. We run the second and call it the first.

Once the rule is maximise shareholder value, the consequences are not moral failures, they are correct play. Care becomes inefficient. Restraint becomes a competitive disadvantage. Long-term thinking becomes a reason to replace you. And the people at the top are inside the same machine: the CEO who picks the long game gets fired, Ben and Jerry's was pushed into a sale, Patagonia's founder had to give the company away to a trust just to escape the market's pull. Good actors, crushed by the rules they play under.

Bad systems beat good humans every time. In our trade, bad systems beat good engineers.

The engineering version is the same shape at smaller scale. The team that invests in maintainability, refuses the dark pattern, keeps the data collection minimal, loses the quarter to the team that shipped the growth metric. Not because they were wrong. Because the scoreboard could not see a single thing they got right.

AI amplifies the capability, not the restraint

This is where the two arguments fuse. Dykstra's distinction is between intelligence and wisdom: intelligence solves the problem in front of you, wisdom asks which problems to solve and when not to act. Wisdom, in almost every definition, has restraint built into it. Wisdom is power plus restraint.

AI is a machine for the first and not the second. It can optimise but it cannot refrain. It can act, soon at scale, but it cannot know when not to. So it amplifies exactly one half of the pair. The gap between what we can do and what we should do, already wide, becomes a chasm, and the historical name for that chasm is catastrophe. Jurassic Park got there first: so preoccupied with whether they could, they never stopped to ask if they should.

The move-fast culture that already shipped the unfinished thing now ships it faster and wider, at machine speed. The one input it never had, "should we", is the one input AI does not supply.

You cannot willpower your way out of an incentive

Both of Dykstra's talks end on a personal practice: make space, pause and get curious, ask what the system is rewarding. Good advice, and not enough, because his own economics argument already explained why. Individual restraint does not survive a system that punishes it. Tell a tired engineer to be wiser and you have restated the vigilance problem we wrote about in human in the loop is theater: you cannot fix a structural gap with more willpower from the person standing in it.

So the pause cannot live in the person. It has to live in the artifact. If restraint shows up on the dashboard only as cost, stop asking people to supply it against the grain. Build it into the product and the business model, so the restrained choice is the default, and the one that survives the market. That is the whole of what we mean by our principles:

• No lock-in. "Stops paying does not stop working." The extraction incentive is removed by construction, because you cannot hold hostage a user you already agreed to release.
• The human on the critical path, not beside it. The machine proposes, a person decides. Restraint becomes a required step in the flow, not a virtue someone has to remember at 3am.
• Self-hosting and sovereignty. The user is not a revenue stream, and their data cannot quietly become the product, because it never leaves their control to begin with.
• Collect less, deny egress by default. The minimal thing is the default and more is opt-in, so the careful choice is the path of least resistance rather than an act of conscience.
• Build for years, not the earnings call. Longevity as an engineering property, few dependencies, a small surface, so maintaining the thing is cheap and there is no incentive to churn it under the customer.
• Confinement for the agent. Capability limits, egress allowlists, a kill switch, an audit trail. The "when not to act" the model cannot supply, supplied by the architecture around it.

None of these are acts of virtue. They are the pause, cast in structure, so that doing the right thing does not depend on anyone being heroic on a bad day.

Do not try to be the good engineer the bad system will beat. Change what the system can see.

You cannot measure restraint, so you cannot manage it into existence, and you cannot exhort it into existence one tired person at a time. You can only design for it: put the pause into the incentive and the architecture, so the restrained choice is also the surviving one. Dykstra says space is the language of the rebellion. For people who build systems, the rebellion is to stop asking humans for the space and to build the space in.