Versions & publishing
Edit an agent safely as a draft, then publish numbered, immutable versions that control exactly what your production agents run.
Agents use a draft-and-publish model so you can edit safely without changing live behavior. The agent you edit is the draft; publishing captures it as a version that production runs use.
Draft vs. published
- Draft — the agent you see and edit. Every change autosaves, but none of it goes live until you publish.
- Version — a numbered, immutable snapshot of the agent's behavior. The active version is what your production agent actually runs.
When your draft differs from the active version, the agent shows an unpublished changes indicator.
What's included in a version
Publishing snapshots the agent's behavior:
| Versioned | Not versioned |
|---|---|
| Persona prompt | Name, avatar, color |
| Tools and app accounts | Triggers |
| Knowledge scope | Attached procedures |
| Tool bindings | |
| Model |
Procedures have their own independent versions, so they're not captured inside the agent's version.
Publishing, reverting, and discarding
- Publish — snapshot the current draft as the next numbered version and make it active.
- Revert — restore the active version's behavior back into the draft.
- Discard — throw away draft changes and return to the active version.
You can also label versions (for example "prod" or "v2.1") and browse version history.
Test before you publish
Run your simulations against the draft and get them passing before you publish. Publishing is what changes live behavior.
Next steps
Models & bring-your-own-key
Choose which AI model powers an agent, override the workspace default per agent, and bill usage to your own provider account with BYOK.
Tool approvals
Add a confirmation gate before sensitive tools run, so a person reviews and approves the agent's proposed action before it executes.