AI Agents have no damping, but they must have one.

Your organisation quietly kills a bad decision every day.

Watch how. An analyst makes a bad call on a Tuesday morning. By Thursday, a colleague has frowned at it, a manager has asked "are you sure?", and the error has died in a meeting that existed for no other purpose.

Nobody designed that. It's just what human organisations do. Every handover, every approval, every sceptical pause at the coffee machine absorbs a little of the error's energy. We call it friction, and we've spent thirty years engineering it away without noticing it was doing a second job.

That second job is damping.

Now replace the analyst with an agent. The bad call doesn't wait for Thursday. It feeds the next decision within seconds, which feeds the next, each iteration built on the last: the agent's outputs becoming its own inputs, at machine speed, with no frowning colleague anywhere in the loop.

Errors in an agentic system don't get absorbed.

They compound.

That's amplification, and it's not a malfunction: it's what any undamped feedback loop does. Engineers have known this since the first bridge shook itself to pieces.

The design consequence almost every agent deployment misses: the friction you're automating away was a load-bearing structure. Human organisations got their damping free, as a by-product of being slow. An agentic organisation gets none. The damping must be engineered in deliberately, and deciding precisely where the absorbing human belongs is an architecture decision, not a staffing one.

Chapter 11 of Architecture of Intellect treats amplification as the first of three systemic agent risks.

Question for the practitioners: in your current agent designs, what actually absorbs an error once it starts moving?

Name the component.