What Is an Agent Application?
In Agent Studio, an agent is not a chat window. It is a governed identity with immutable versions. An application is how that identity shows up in the world — branded, published, and constrained.
The factory metaphor
Draft config → Submit → Approve → Provision → Publish → Serve sessions
│ │ │ │ │ │
builder UI approval version runtime channel Agent Gateway
record adapter
- Define — instructions, model, tools, budgets, knowledge pointers.
- Govern — approval request, SoD, audit event, immutable version.
- Provision — worker asks a runtime adapter (Claude / local) for a deployment.
- Publish — bind that version to hosted web, embed, API, and/or desktop.
- Serve — end users talk only to the Agent Gateway, never to provider keys.
Objects you will meet
| Object | Role |
|---|---|
| Agent definition | Named identity inside an organization |
| Agent version | Immutable snapshot of config |
| Approval request / decision | Human gate with SoD |
| Deployment | Runtime environment for an approved version |
| Application | Product shell (branding, template, slug) |
| Publication | Channel-specific release (hosted_web, embed, api, desktop) |
| Publication token | Runtime credential for embed/API (pub_…) — not a login |
Why “application” matters for AI
Shipping a raw agent chat to customers skips product concerns: origin allowlists, CSP frame-ancestors, postMessage token delivery, channel-scoped AuthZ, and spend enforcement. Agent Studio makes those publication properties, not after-the-fact middleware hacks.
[!TIP] Think of Claude (or any future adapter) as the engine room. Agent Studio is the ship: bridge (control plane), cargo rules (governance), and hatches (publish channels).