An agent prepares a change. The user edits it in an interactive interface. Then the agent continues—but which version of the world is it acting on?
MCP Apps bring people and agents into the same workflow. Making that collaboration work requires deliberate choices about shared state, context, approvals, and who acts next.
Through one worked example, this talk examines how to keep tool responses and interfaces consistent, handle human edits while an agent is working, make approvals refer to concrete changes, and return useful context when control passes back. Attendees will leave with practical interaction patterns and failure scenarios for building MCP apps that people and agents can use together.
