Reintroducing friction isn’t a bug. It’s a political choice.

A costly choice: it slows things down, complicates them. Rarely made — unless you understand that the cost of friction is lower than the cost of fragility.

Some propositions

1) Preserve non-automated zones: Not out of nostalgia, but for cognitive hygiene. Spaces where you keep doing things yourself, making mistakes, understanding why things break. The point is that it’s a choice of competence, not technological lag.

2) Organize cognitive fire drills: Regularly unplug the agent and replay by hand. Verify you still know how. It’s tedious, it’s slow, but that’s exactly why it’s useful.

3) Establish a right to incomprehension: Not a right to understand everything, that’s impossible. But a right to demand justification before approving. “The agent says…” cannot be a sufficient argument.

4) Evaluate agents on recoverability, not results: A good tool is one you can unplug. If the agent goes down tomorrow, how long before the collective is operational? If the answer is “several weeks,” there’s a design problem.

5) Document the trade-offs, not just the outputs: When you configure an agent, you make choices. Those choices should be traceable, explicit, contestable and not buried in a prompt no one rereads.

Slow-down could be the price

These measures slow things down. They require unproductive time that is, political time. That’s precisely why they won’t be taken by default.

Friction isn’t an obstacle to efficiency. It’s the price of resilience.