Agents and deployments¶
An Agent in Mission Control is managed application content: instructions, a model, tools and account choices. Saving and publishing that content gives you versioned applications that people can use in Chat or invoke from code.
Choose where your agent runs¶
| External application | Managed Agent | |
|---|---|---|
| Code and execution | Your process and agent framework | Mission Control's managed runtime |
| Configuration | Your application code | Agent definition in the Console |
| Authentication | User or ordinary service-account virtual key | Caller access plus internal managed execution credentials |
| Release process | Your application deployment | Saved snapshot and reviewed publication |
| Starting point | Run your own agent | Build your first Agent |
External code does not need a Builder Agent record. Both paths use the governed model and tool gateways.
Draft, snapshot, Preview and Live¶
The draft is the definition you are editing. Save snapshot records an immutable version of content and account defaults. Private Preview executes a saved snapshot using your developer access and limits. It can consume model and tool resources.
Review & publish checks the proposed version, execution access, accounts, funding and sharing. Publication pins a snapshot to a deployment, the destination consumers invoke. The Live tab shows that deployed version. Editing a draft does not change live behavior until you publish again.
One Agent can have several deployments. Choose the intended destination when publishing; adding a deployment is an explicit action. Restoring an older version retains current access, accounts and caps subject to compatibility checks. It is not a rollback of all administrative state.
Caller access and execution access¶
Two decisions matter when a published Agent runs:
- Who may invoke it? Chat uses the deployment's Chat audience and the actual signed-in person. API access uses explicit allowed callers plus an eligible virtual key with
deployment.invokefor that deployment. - What may it do? Its model and tool calls use managed execution access, bounded by Team policy and funded by the publisher Team. Automatic access follows the saved Agent choices; an administrator can impose a custom restriction.
An invoke-only caller can use a published Agent without direct permission to call its models or tools. Conversely, a developer who can call a model does not automatically have access to every published Agent.
Sharing in Chat does not grant Team administration or observation access. Private conversations remain owned by their actual user and membership. A Session ID groups activity; it does not transfer conversation ownership.
Preview, evaluation and a production test¶
Preview helps you iterate on a saved draft. An experiment measures a pinned subject against a versioned dataset. Test live version checks the actual deployment with live accounts and Team funding. Each serves a different purpose; a successful Preview does not prove the production account works.
Follow Build, Preview, and Publish, Live administration, Chat, or deployment invocation.