An LLM answers. An agent acts.

You assessed the model. Nobody assessed the agent.

That sentence sounds wrong. Your risk committee has a file: benchmarks, vendor due diligence, red-team results. But everything in that file describes the model. And the model is not the thing that acts in your enterprise.

An LLM is a timeless function: a mapping from context to tokens, no before and no after. Each call is an isolated event, like evaluating sin(x). It doesn't forget yesterday's conversation; there was no continuity to forget. Between calls, there is no one there. Ask 1,000 times, get 1,000 independent evaluations. That's the whole machine.

An AI agent is a different kind of thing entirely. It's an actor: a process that runs in time and acts on the world, its Observe→Act loop passing through your real operations, not through the model's insides. Its outputs become its own future inputs. Strip away the sequence and the world-coupling, and the agent vanishes, leaving a series of unrelated calls to stateless functions.

Not a difference of degree. A difference of category.

And the category carries the risk: a function cannot amplify anything. Amplification, damping, resonance, instability are properties of systems that live in time and feed back on the world. A market has them. An ecosystem has them. A control circuit has them. A function doesn't. Wrap an LLM in a loop, connect it to your operations, and it joins the class of systems that live in time and feed back on the world. Every risk framework you own was written for functions.

Which returns us to the first sentence. The model was assessed. The vendor was assessed. The actor that runs in your enterprise hour after hour was assessed by nobody, because no methodology on your risk register knows it exists. And what that methodology would have to examine is not the actor at rest but its trajectory: the path of its actions through your firm. Testing the actor tells you what it can do. Only the trajectory shows what it does.

Chapter 11 of Architecture of Intellect is built on this distinction.

A falsifiable claim to argue with: no amount of testing the model tells you how the agent will behave, for the same reason no amount of testing a neuron tells you what the brain will do. Prove me wrong in the comments. I'll answer every serious attempt.