Agent Applications - Agent Applications

ChatGPT and Claude use agent widgets. An OpenClaw agent may appear as a WhatsApp or Telegram contact. GitHub Copilot’s coding agent and Cursor Cloud Agents can start from an issue or pull request and return their work there. A Microsoft Copilot Studio autonomous agent can wake in response to a business event without presenting a conversational interface at all. These systems have different interfaces and jobs. Each gives a tool-using agent a continuing scope of work. Later messages or events return to that work, and the results remain after a run ends. The paper calls that product an Agent Application: one or more persistent, tool-using AI agents that produce or maintain durable results. The application sets tools, permissions, and other guardrails. A request, schedule, or outside event starts a run. The agent reads the current state and chooses the next action. The results outlive the model call that produced them.

Four properties

A transcript can remind an agent what it said. A workspace lets the agent continue the work. Durable work is a result that remains after the run that created it: an artifact, a workspace change, a record in another system, or a process still running. A file the agent wrote can be both. The privacy domain is who that instance belongs to: a person, a project, a customer, or a company. A chat interface does not make a system an Agent Application. The same agent may appear as a full-screen conversation, a copilot inside another product, a phone number, or a background worker. These systems fall outside the category:

The boundary is fuzzy, as it is between a web page and a web application.

Products that already have this shape

If this table is missing a product, or includes one it should not, open a discussion.

First review

Take one product you know well. Using the four properties, is it an Agent Application, not one, or a boundary case? If the definition is too tight, too vague, or missing something you need in order to make that call, say what. A missing facet only counts if you need it to tell Agent Applications from chatbots, workflows, or task runners. Post in discussion 1, or email amol@agentapplication.io.

After desktop, web, and mobile

Desktop, web, and mobile applications each gave developers a recognizable way to assemble a complete product. Agent Applications change how control, context, and work are handled.

Instances diverge

Long-lived instances accumulate different facts, artifacts, generated code, and unfinished work. If they can also add or edit natural-language instructions, they diverge in program as well as state. Operators then manage a fleet whose members no longer share one program. Rollout, audit, and policy have to follow each instance’s local program and state, not only the shared release. A new release can improve the shared application while failing against the state of one instance. If the instance edited a published instruction, the upgrade merges that instruction by meaning. Copy and line-based merge miss how one instruction changes another. The Hello World example is the smallest complete case: a notebook agent saves a note, suspends, and uses that note to update a briefing on a later run.

Framework, platform, and store

The layers describe responsibilities, not seven products that a team must buy. One provider may combine several layers. An Agent Application Framework is a harness that also supplies a coherent project structure, development tools, evaluations, packaging, deployment, and production runtime support. An Agent Application Platform turns a tested project into an immutable release, provisions long-lived instances with durable workspaces, and enforces identity, permissions, spending limits, network destinations, and approvals. An Agent Application Store is a catalog and commerce service through which people find, acquire, and pay for Agent Applications. Keeping those responsibilities distinct makes architectures and vendor claims easier to compare. Each framework keeps its own project and release formats. MCP, Agent Skills, Agent Plugins, and A2A already cover some of the boundaries where independent systems meet. The paper is the architecture: the definition, the capability model, the lifecycle, the security and governance view, and the open questions. It includes guided reading paths for application developers, framework authors, platform operators, buyers, and standards contributors.

Contribute

The paper is a working draft. Open threads:

• Which current products are Agent Applications?
• Is Dependabot an Agent Application?
• How should a platform upgrade instances that edited their own instructions?

Use Discussions for other proposals, architecture questions, boundary cases, and prior art. Open an issue for a concrete error, missing source, or broken page. Send a pull request for a focused correction, example, or documentation improvement. The first review is one product you know, then the definition. CONTRIBUTING.md explains how to prepare a pull request.

Cite

Correspondence: hello@amolkelkar.me