Agent-Integrated Software: Interaction Contracts and Continuous Assurance

TL;DR - This perspective paper proposes Agent-Integrated Software, a pattern for safely embedding agents into applications where users can revise goals or shared objects during execution. It argues that explicit interaction contracts and continuous assurance are needed to make such integrations dependable and maintainable.

• Introduces Intent-Level Interaction Abstraction to represent how users inspect, modify, and control delegated tasks.
• Models agent-integrated applications as open transition systems connecting task-level events to application behavior.
• Defines contracts covering task bindings, role-based authority, control transitions, and evidence of outcomes.
• Separates admission invariants from agent planning and outlines research needs in controlled execution, evaluation, development tooling, and human supervision.