Designing the disagree path
Why human override is the most overlooked feature in modern AI tooling.
Section · 02 / AI + Systems
Designing for intelligent systems — and the humans steering them.
A growing body of work on LLM copilots, agentic workflows, and the operational surfaces that wrap them. The hard problems aren't model outputs — they're trust, control, recoverability, and giving people a way to disagree.
Principles
Notes on designing for AI.
-
01
Make the model legible.
Surface what the system saw, what it inferred, and what it's about to do. Confidence is a UX problem, not just a metric.
-
02
Disagreement is a first-class interaction.
If users can't override, correct, or back out — the system isn't a tool, it's a verdict. Build the disagree path before the happy path.
-
03
Agentic ≠ autonomous.
Agents should ask for permission at the right altitude — never silent on consequence, never noisy on routine.
-
04
Speed is a feature; predictability is the product.
A fast, unpredictable assistant is unusable in operations. Design the rails, then let the model run.
Talks & writing
Talks and writing are placeholders for now — drop links in and they'll click through.
Working on something AI-shaped?