Agent Applications

A working paper · Draft 0.7 · August 2026 · Read the full paper · Hello World · Cite

An Agent Application is a software system whose primary unit of execution is one or more persistent, tool-using agents operating in durable workspaces to accomplish work over time and produce or maintain durable artifacts.

Desktop, web, and mobile software each developed its own application model. Agent software now needs one too. An application that completes work over time needs somewhere to keep that work. This affects how the software is written and packaged, how its instances are created and governed, and how the product is priced.

The building blocks already exist

By 1994, web applications were beginning to emerge from a few basic pieces. HTTP moved requests, HTML described pages, and browsers displayed them. JavaScript later added behavior. Frameworks made the work repeatable and helped settle what developers meant by a “web application.” Agent developers now have comparable building blocks: the Model Context Protocol for tools, portable Agent Skills, sandboxed compute, and harnesses that run the agent loop. Today’s harnesses are early versions of Agent Application frameworks. As they add project structure, development tools, evaluations, packaging, and deployment, they make agent development repeatable.

Four defining properties

Property	What it looks like
Persistent agent instance	A new message or event returns to the same agent, privacy domain, and ongoing work
Agent-directed control flow	The agent reads the current state and chooses its next action within application guardrails
Durable computational workspace	Files, instructions, memory, code, and other working state remain available to later runs
Durable work	Artifacts, workspace state, external records, or continuing processes outlive the event that created them

A workspace defines a logical boundary for durable state. A platform may store that state in databases, object storage, or retained sessions. In the canonical design, each instance gets an isolated persistent virtual filesystem that the platform restores on demand. The agent can read everything in its workspace. The people, records, and work that can safely share that workspace form its privacy domain. ChatGPT, Claude, Gemini, Lovable, Replit Agent, OpenClaw, and Lightfield already organize ongoing work this way. Depending on the product, an instance belongs to a user, project, personal agent, company workspace, or customer relationship. The agent widget is the most common interface for an Agent Application. It may fill the screen, sit beside another application, or open in a popup. It handles authentication and session switching. Users can choose models and harness options, type or talk, upload files, and control the running agent. The conversation view shows messages alongside progress, reasoning summaries, tool calls, and approvals. It also reports errors and other runtime events. The widget renders artifacts and interactive UI returned by the agent. OpenAI ChatKit, CopilotKit, and AI SDK Elements package many of these parts. AG-UI defines the event stream between an agent and its frontend. MCP Apps lets tools place forms, dashboards, viewers, and other interactive controls inside the conversation.

Application models by era

Era	Application model	What persisted
Personal computing	Desktop application	Local files, installers, the OS
Internet	Web application	Servers, databases, the URL
Mobile	Mobile application	Device identity, sensors, the store
Agent computing	Agent Application	Instances, workspaces, durable work, lineage

Built as a project, deployed as instances

A developer builds an agent project in the native layout of an agent harness or Agent Application framework, then publishes it as an immutable release. The project declares its runtime requirements. An Agent Application Platform runs it in its native form instead of translating every system into a lowest-common-denominator schema.

account-operator/
|-- <harness- or framework-native entrypoint>
|-- skills/
|   |-- account-review/
|   `-- renewal-planning/
|-- agents/
|   |-- researcher/
|   `-- reviewer/
|-- tools/
|-- policies/
|   |-- approvals.md
|   `-- external-actions.md
|-- knowledge/
|-- artifacts/
|-- evals/
|   |-- behavior/
|   |-- security/
|   `-- regression/
|-- migrations/
`-- app.yaml

An application may create an instance for a user, team, customer, project, ticket, or another continuing relationship. Each instance accumulates its own workspace, artifacts, instructions, and history. Instances diverge as they work, so updating a fleet involves more than copying the same release over each one.

Cite

@misc{kelkar2026agentapplications,
  title  = {Agent Applications: A New Software Architecture for the AI Era},
  author = {Amol Kelkar},
  year   = {2026},
  note   = {Working Draft 0.7},
  url    = {https://agentapplication.io/paper}
}

Written by Amol Kelkar. Any platform may implement the definitions and portability contracts.

Overview

Concepts

About the paper

The building blocks already exist

Four defining properties

The agent widget

Application models by era

Built as a project, deployed as instances

Cite

​The building blocks already exist

​Four defining properties

​The agent widget

​Application models by era

​Built as a project, deployed as instances

​Cite

The building blocks already exist

Four defining properties

The agent widget

Application models by era

Built as a project, deployed as instances

Cite