Introducing Versioning

ElevenLabs Versioning applies software release discipline to the voice and chat agents companies build on its platform. With Versioning, each saved change produces a new version that records the whole configuration: prompts, voice, workflows, tools and the underlying language model. The announcement says diffs and editor attribution show what changed, who changed it and when, so a team that sees an agent start answering differently can trace the change to a specific edit and check whether anyone approved it.

The second part is a staged rollout. A team can send a defined share of conversations to a new version while the rest keep going to the stable one, and compare the two before switching everyone over. The versioning docs describe how this works. Versions are immutable snapshots with their own IDs. They sit on named branches, much like git branches, with a main branch that cannot be deleted. Traffic percentages across branches must add up to exactly 100, and conversations are routed deterministically by conversation ID.

ElevenLabs also pitches the feature at compliance, risk and legal staff. Because a fixed configuration lies behind every conversation, the company says an organization can point to the exact setup that handled a given call instead of reconstructing it after the fact. One practical constraint is in the docs rather than the post: versioning is opt-in per agent and cannot be turned off once enabled. ElevenLabs later built its Experiments A/B testing feature on top of this system, and the Experiments docs say versioning has to be enabled before an experiment can run.