Agent-native applications: a new application paradigm
Agent-Native Applications: A New Application Paradigm
AI summary
The paper proposes a new paradigm, "agent-native applications": LLM agents take on the application's core logic and, after authorization, can adjust the application graph at runtime. The architecture represents the application as a portable graph of agents, tools, data sources, and human-in-the-loop checkpoints, and constrains authorized behavior with a behavioral envelope in a declarative specification. At runtime, the application orchestrator serves as the control plane and the agent mesh as the data plane and generates audit events; an executable ANR prototype demonstrates some of the core mechanisms.