Need to roll back a voice agent change fast? How versioning and rollback work on Vapi, Retell and ElevenLabs
This free guide is for teams running voice agents on Vapi, Retell or ElevenLabs. For each platform it lists what is versioned natively, what you must script yourself, and the steps to go back.
Checked against the Vapi, Retell and ElevenLabs docs on 18 November 2026.
Vapi: draft, publish, history and restore
Vapi versions assistants and tools natively, and restoring an assistant version takes effect immediately. The week of 17 August 2026 is when the docs date the feature.
- Native
- Drafts are per user and do not affect live calls. Publishing creates a numbered version that becomes current. Restoring creates a new version from the old configuration and makes it current. For example, restoring v1 while v2 is current creates v3. Restoring replaces any current draft. Tools restore into a draft and need a publish.
- Rollback steps
- Restore the earlier assistant version in the dashboard. For a canary, set the bad version to 0 percent in the traffic split. Squads keep version history, but the docs say the dashboard has no control to make an earlier squad version current. The way back is to update the squad with the old member list.
- You script
- Saving the JSON of every assistant before a fleet change. The docs describe export in the dashboard but no API route for restore, so a scripted rollback means reading the saved configuration and updating the assistant with it. The docs do not spell that out, so test it on one assistant first.
- Config as code
- Vapi GitOps saves a snapshot on every deploy and has list and restore-to-timestamp commands. The docs do not say whether a rollback creates new Vapi versions or overwrites the live configuration.
Retell: versions, tags and the deleted LLM trap
Retell rolls back by version: create a draft from any past version, or move a tag or number binding back. A deleted Retell LLM is the one thing you cannot roll back.
- Native
- Numbered versions starting at V0, drafts you can edit, and published versions that are read-only. Each agent starts with prod and staging tags and can have up to 10 tags, according to the docs. A page for comparing agent versions side by side exists.
- Rollback steps
- The docs example moves the prod tag to V3 and back to V2 to return to the previous configuration. Or create a draft from the earlier version and publish it. If a version has phone numbers or tags attached, remove those assignments before you delete it.
- You script
- Keeping your own copy of prompts and LLM config. Retell staff on the Retell community forum recommend keeping prompts and configs in your own version control and using the update call instead of creating and deleting.
- The deleted LLM trap
- Retell staff on the Retell community forum say a deleted Retell LLM configuration cannot be recovered and that no backup or export is kept. The delete call refuses by default if an agent still uses the LLM, but with force set it deletes anyway and leaves those agents pointing at an LLM that no longer exists.
- Shared subflows
- The docs we read describe no version history for library subflows, so an edit to one has no documented way back. Keep a copy before you change it.
ElevenLabs: immutable versions, branches and deployments
ElevenLabs keeps full version history on branches. The versioning page has no revert command, so the practical rollback is to send traffic back to Main or read an older version.
- Native
- Versions are created when you save to a branch and cannot be modified afterwards. Branches work like git branches, with an undeletable Main. Deployments split conversations across branches by percentage, and the percentages must total exactly 100. The experiments page says full version history is preserved, enabling rollbacks.
- Rollback steps
- Deploy 100 percent of traffic back to Main, or retrieve the agent at an earlier version or branch tip by its ID. The versioning page does not mention restore or revert, so confirm the exact step on one agent first.
- You script
- Recording the version each number and deployment used before a change. Rolling an agent back does not undo an edit to a shared tool, because the earlier version references the same tool ID.
What to script on every platform
Native version history covers one agent at a time. A fleet needs a saved snapshot, a written rollback runbook and a read-back after the rollback.
Snapshot before every change
Save each agent's config and the version each phone number points at, in git.
Write the rollback as a runbook
Name the pointer to move, who may move it and how you verify the result. Do it before the change, not during the incident.
Roll back shared parts separately
Tools, subflows and shared LLMs are not undone by rolling an agent back. Keep their own copies.
Read back after the rollback
Confirm each agent and number points where the runbook says, and check the version on a fresh call.
Sources
Official documentation and forum pages this guide relies on, as read on 18 November 2026. Platforms change fast, so check the page before you act.
- Vapi docs: assistant versions, restore and export
- Vapi docs: tool versions
- Vapi docs: versioning with squads
- Vapi docs: traffic splitting
- Vapi docs: config as code
- Retell docs: agent versions and tags
- Retell API reference: delete Retell LLM
- Retell docs: conversation flow subflows
- Retell community forum: deleted LLM configuration (topic 2349)
- ElevenLabs docs: agent versioning
- ElevenLabs docs: experiments
- ElevenLabs docs: editing tools