Agent Applications

A working paper · Draft 0.9 · August 2026 · Read the paper · Review the draft · Cite

Agent Applications are AI systems built around persistent, tool-using agents. A request, event, or schedule can start work. The agent chooses intermediate actions within application guardrails, while its instance retains a workspace, authority, artifacts, and unfinished work after the run ends. The working paper presents a vendor-neutral reference architecture for building and operating these systems. It identifies the main elements, assigns responsibilities to them, and follows an Agent Project through release, deployment, operation, and change. It also locates standards such as the Model Context Protocol (MCP) and Agent Skills at the boundaries they cover. The paper is a working draft. We invite readers to test its claims against real products, implementation experience, prior art, and counterexamples.

##

​

What the paper gives you

• A common vocabulary for projects, releases, instances, workspaces, artifacts, frameworks, platforms, and stores
• A build-and-operate method for turning a use case into a tested application design
• A capability model for comparing frameworks and assigning responsibility
• Lifecycle, security, and governance views for operating long-lived instances without losing their accumulated work
• An account of existing standards and the exchange boundaries where future contracts may help
The paper includes guided reading paths for application developers, framework authors, platform operators, buyers, and standards contributors.

##

​

Core terms

| Term | Meaning | | ----------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Agent Application | The complete product built around one or more persistent, tool-using agents that produce durable work. | | Agent Project | The framework-native project representation used to create releases. | | Agent Application Framework | A harness that includes coherent project structure, development tools, evaluations, packaging, deployment, and production runtime support. | | Agent Application Programming | The discipline of decomposing a use case across code, natural-language programs, tools, agents, workflows, state, and policy. | | Agent Application Platform | A service that turns tested Agent Projects into immutable releases and operates long-lived instances with durable workspaces, identity management, security controls, governance, observability, evaluation, and fleet management. | | Agent Application Store | A catalog and commerce service through which people find, acquire, and pay for Agent Applications. |

##

​

Read the work

• The paper presents the full argument, category definition, and architecture.
• Hello World shows the model in a small application.
• About this paper explains the audience, scope, and status of the work.
• Companion documents lists the practical and technical documents planned around the paper.

##

​

Contribute

We welcome community engagement as the paper evolves.

• Use GitHub Discussions for open-ended proposals, architecture questions, and boundary cases.
• Open an issue for a concrete error, missing source, or broken page.
• Send a pull request for a focused correction, example, or documentation improvement.

##

​

Cite

@misc{kelkar2026agentapplications,
  title  = {Agent Applications: A Reference Architecture for Building and Operating AI Agent Systems},
  author = {Amol Kelkar},
  year   = {2026},
  url    = {https://agentapplication.io/paper}
}

Correspondence: hello@amolkelkar.me

Suggest editsRaise issue

⌘I