Create Agent Versions

An agent version is a snapshot of the agent’s settings and configuration. Live versions and versions attached to an experiment are immutable. An unattached, non-live candidate can be edited before review. Experiments compare saved versions; they do not compare unsaved Draft changes.

Plan the versions for a test

Use one version as the control and create another version with the single change you want to measure.

Keep unrelated changes out of both versions. If the treatment changes the greeting, do not also change escalation logic or knowledge sources in the same test.

Create a version without making it live

1. Make the intended change to the agent.
2. Select Deploy in the application header.
3. In New revision, review Changes and add a description that names the behavior you changed.
4. Clear Deploy revision to live agent if you only want to save the version for an experiment.
5. Select Create Revision.
6. Open Deploy -> Versions and record the version number and description.

For a control, you can use the current live version or repeat these steps to save a separate baseline. Create every version before you open the experiment setup.

Creating a revision with Deploy revision to live agent selected changes the agent’s default live version immediately. It is not an experiment rollout.

Create a version from JSON

Applied CLI can create or update a candidate from the complete version JSON:

applied agents <agent_id> revisions get <live_revision_id> -f json
applied agents <agent_id> revisions create --body-file candidate-revision.json -f json
applied agents <agent_id> revisions <candidate_revision_id> update \
  --body-file candidate-revision.json -f json

Copy the full data object returned by get, then change only the intended field. The API rejects non-object data, live-version edits, and edits after a version is attached to any experiment. Create a new candidate instead of rewriting experiment history.

Choose a version in an experiment

When you create an experiment, every variant must point to a saved version. Use distinct version configurations for the control and each treatment, and give each variant a short name that explains its role. Two version IDs with identical JSON are not valid experiment arms.

Recommended descriptions:

• Control — current refund flow
• Treatment — ask for order ID first
• Treatment B — shorter opening message

Avoid descriptions such as test, new version, or latest; they are hard to interpret after the experiment ends.

Change the default live version

Open Deploy -> Versions, find the version, and select its rocket action. Confirming the action makes that revision the default live version.

Traffic rules are evaluated before the default fallback. A rule that targets a specific version or a running experiment can still override the live version for matching traffic.

Roll back

A rollback is another deployment action:

1. Open Deploy -> Versions.
2. Find the last known-good version.
3. Select its rocket action and confirm Deploy.
4. Review Traffic rules for any rule that explicitly targets a different version or an experiment.

Deploying an older version does not rewrite or delete newer versions.

Next step

Continue to Create and Run an Experiment.