Agent versions
Iterate on a live agent safely using drafts, publishing, and rollback.
2 min read
Agents use draft-and-publish versioning so you can iterate on an agent without affecting the users talking to it right now.
Draft vs. live
Every agent has two states at any time:
- Live version — the one currently serving requests. This is what users see when they pick the agent in chat and what webhooks and delegations call.
- Draft version — your work in progress. Editing an agent's instructions, knowledge, or tools updates the draft. Nothing users see changes until you publish.
The top-right corner of the agent editor shows which version you're viewing — Draft or Live — and lets you switch between them.
Publishing a draft
When you're happy with the draft, click Publish. Publishing:
- Records the previous live version in the version history.
- Makes the draft the new live version.
- Clears the draft state. Future edits will start a fresh draft.
Any conversation that was mid-reply when you published continues to completion using its original version — nobody sees a mid-turn personality change.
Version history
The version-history dialog shows every published version of the agent, with the author, publish time, and a brief summary of what changed. For each past version you can:
- Compare — diff its instructions against the current live version.
- Restore — make it the new draft, which you can then publish.
Rollback
If a published change causes problems — wrong tone, bad answers, broken tool access — open version history, pick the last known-good version, and click Restore then Publish. The rollback is immediate for all new conversations.
File-based agents
Agents defined in TALE_CONFIG_DIR/agents/*.json don't use the UI versioning system — their history is whatever your git repository records. See AI-assisted development for the file-based workflow.