Welcome to Agent Studio
Agent Studio is an enterprise Agent Application Factory: define an AI agent once, govern it centrally, and publish it as a secure web app, desktop shell, embedded copilot, or API — without handing end users the keys to your model stack.
Define once · Govern centrally · Publish anywhere · Provider-neutral core
Most "agent platforms" stop at a chat box. Agent Studio treats the agent as a product: versioned configuration, human approval gates, runtime provisioning, multi-channel publication, and a hardened Agent Gateway that sits between every surface and the model provider.
Why this exists
Frontier models are powerful and still dangerously easy to ship wrongly:
- A prompt change goes live with no review trail.
- An embed leaks tokens in a query string.
- A “demo adapter” quietly runs in production.
- Spend, concurrency, and tool allowlists are afterthoughts.
Agent Studio is built so the default path is the governed path. Claude Managed Agents is the first execution provider. The domain stays provider-neutral through an AgentRuntimeAdapter boundary — so tomorrow’s runtime does not force you to rewrite tenancy, approvals, or publications.
What you get
| Capability | What it means for an AI product |
|---|---|
| Immutable versions | An approved agent identity is a snapshot, not a mutable chat config. |
| Separation of duties | Builders submit; approvers decide. Self-approve is off by default. |
| Agent Gateway | Every session authenticates, authorizes, budgets, and audits through one door. |
| Publish channels | Hosted web, embed, API, and desktop — same approved version. |
| Enterprise identity | Email/password, optional OIDC SSO, TOTP MFA, privileged MFA gate. |
| Ops package | Production Compose, Caddy TLS, backups, retention purge, AuthZ smokes. |
[!TIP] New here? Start with Getting Started, then ship something real with Ship your first governed agent.
Who this handbook is for
- Builders shaping agent instructions, tools, and application branding
- Approvers / platform owners enforcing governance and spend
- Operators deploying the production Compose stack
- Security engineers reviewing identity, AuthZ, and threat boundaries
- AI architects evaluating how a factory model maps onto Managed Agents
Reading this book
Use the sidebar like a factory floor map: Overview → Guides → Product → Developing → Security → Operations. Search lives in the header. Use the sun/moon control for light/dark theme.
[!NOTE] Deep engineering notes also live in the repo under
docs/. This handbook is the narrative, AI-centric surface; the repo docs remain the source of truth for ADRs and gap analysis.