Designing for People
The interface around the agent.
Two companies launch the same agent, with the same model. Customers love one, and don't trust the other. The difference is the interface. The first shows what it's doing, asks before it acts, and makes mistakes easy to undo. A working agent isn't enough. People need to feel in control of it.
- Same agent, different reactions
- The interface decides
- People need to feel in control
Principles of agent design
Show what it's doing, step by step. Show the plan before it starts on anything large. Let people interrupt it at any moment. Make every action reversible wherever you can. Show sources, so that people can check. And be honest about what it can't do. Trust is built by predictability, rather than by clever replies.
Good and bad approval prompts
Look at how an approval request should read. The bad one asks, approve action? People just click yes. The good one says what will happen, why, what it'll cost, and how to undo it. A person can make a real decision in two seconds.
Bad
- Approve action?
- refund(order_id=A17)
- Yes or no
Good
- Refund $640 to Priya
- Order A17 arrived damaged
- Can be reversed for 24 hours
Calibrating trust
You want people to trust the agent exactly as much as it deserves. Too little, and they'll check everything, and save no time. Too much, and they'll stop looking, and miss the mistakes. So show confidence when you can, make sure that early experiences go well, and be open when it fails. Start on low risk tasks, and let people grow into it.
Summary
Show the work, ask clearly, and make it reversible. Next, a type of agent where every millisecond counts.
Key takeaways- Design for control
- Approvals must be readable
- Calibrate trust