Skip to main content
Agent Project is the harness- or framework-native source directory a developer commonly calls an “agent” or “Agent Application.” It contains the code, natural-language programs, declarations, policies, dependencies, and tests used to create releases.
An agent project keeps the layout prescribed by its harness or 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. A full Agent Application framework connects the project layout to local development, evaluation, packaging, deployment, and the production runtime. The project also declares what it needs from an Agent Application Platform, including models, tools, workspace features, permissions, schedules, evaluations, artifact support, and upgrade rules. Publishing fixes the source and its tested runtime requirements in an immutable release. Developers will usually say they are “building an agent.” The more precise term agent project matters when the source directory has to be distinguished from an instance, a session, or the runtime agent doing the work. See the paper, sections 7, 9, and 10. Appendix A contains a candidate project manifest.