If you're designing context for agents, you're designing it for humans too, whether you realize it or not
I was listening to Stefanie Hutka on Finding Our Way and it got me thinking about my own work on context design. If you're designing context for agents, you're also designing the context of the humans who live alongside them, whether you realize it or not. So you'd better work on both at the same time.
She talks about values versus vision.
Stefanie Hutka, Finding Our Way: "vision is dependent on seeing the future"
And nobody can see very far right now. What the models can do changes every few months, so whatever vision you had for your product is out of date pretty fast. Values don't have that problem. If you decide the expert should stay the expert, that still holds after the next model release.
Her book is called What Your Machines Should Do. That's the same coin as what the humans should do. If you're thinking about what your machines should do, you're also thinking about what the humans are doing. I'd say that's the work that uses their expertise, that they find fulfilling, that connects them to other people.
That's where context comes in. When you design an agent's context, you decide what it knows, what it does on its own, and what it hands back to a person. What it hands back is the context the human has to work with. You made that decision even if you only thought about the agent. If all the person gets is an approve button and no reasoning, they become a liability sink: they're accountable, but they have no way to judge what they're approving.
She tells a good story about Toyota. Before cars they made looms, and the founder built one that stops itself when the thread breaks, so nobody has to stand there watching it. The machine knows its limits and hands back to the person.
So context design is two jobs. The context for agents is the engineering side: what the agent needs to be good at its work. The context for humans is a lot of social science: how people work, what they need to know to make a call, what keeps the work theirs. Do the first without the second and you've still designed the human's context. You just did it by accident.
The episode is Automate All the Things!? A UX Reality Check on AI on Finding Our Way with Peter Merholz and Jesse James Garrett. Hutka writes Sendfull, and her book What Your Machines Should Do is out in November from Rosenfeld.