Skip to main content
An Agent Project is the native project representation for a harness / Agent Application framework, usually a source directory. It contains the code, natural-language programs, declarations, knowledge, dependencies, and tests used to create releases.
An agent project keeps the layout prescribed by its harness / Agent Application framework. A Claude Code project, for example, may use CLAUDE.md, .claude/, and .mcp.json. Other frameworks have their own files and conventions. The project stays in that native form when it moves from local development to production. Agent Projects do not share a standard directory structure. They may contain natural-language instructions, computer-code tools, system integrations, evaluations, package dependencies, and platform configuration. Section 12 groups these parts so developers can compare frameworks. Each framework retains its own project schema, and most projects will use only part of the model. An Agent Application framework ties the project to the rest of the development process. It defines how developers run and test the project locally, package it, deploy it, and use it in production. It may also tell an Agent Application Platform which models, tools, workspace features, permissions, schedules, artifact support, and upgrade behavior the project needs. Plugins may be declared as project dependencies. They can add skills, MCP servers, hooks, or other components. Some applications also allow a runtime agent to install a plugin as an instance-local program change. Each framework decides how to install and record plugins. When a publisher creates a release, the framework and platform record the project source, resolved dependencies, runtime, active components, requested authority, evaluations, and migration rules in their native release format. Developers will usually say they are “building an agent.” The term agent project becomes useful when the source directory must be distinguished from an instance, a session, or the runtime agent doing the work. See the paper’s vocabulary, capability model, and application lifetimes. Appendix A shows one illustrative project layout.