Skip to main content
Agent Application Framework is an agent harness that includes coherent project structure, development tools, evaluations, packaging, deployment, and production runtime support.
An agent harness runs the loop around a model. It assembles context, calls tools, and loads skills or subagents. An Agent Application Framework extends that runtime into development and production. It tells developers how to structure a project, test it locally, package it, and run it in production. Section 12 of the paper uses a capability model to compare frameworks that expose different APIs and arrange projects differently. It covers agent programs, instructions, tools, workflows, state, interfaces, artifacts, triggers, policy, evaluations, and lifecycle operations. A framework can provide these capabilities through code, configuration, hosted services, or any combination of them. The choice of framework affects the application. Frameworks differ in their model support, context management, tool systems, permissions, delegation, checkpointing, and debugging. Those choices affect behavior, portability, and cost. Several standards connect systems at specific boundaries. MCP connects hosts to tools and data. Agent Skills packages reusable instructions, and Agent Plugins packages skills with MCP configuration. AG-UI connects a backend to a frontend. A2A connects independently deployed agents. Each framework retains its own project and release formats. An Agent Project keeps the native layout of its framework. An Agent Application Platform can support several frameworks without translating every project into one generic schema. See the paper’s stack, vocabulary, build-and-operate method, and capability model.