Use this worksheet to decide whether one existing agent answer follows an approved source. BIG CHANGE did not enter an account or run an agent.
The big change
- What changed: Next Reply checks one answer, not a whole conversation or tool invocation.
- Why it matters: The owner can compare the answer with the exact source it should follow.
- What to watch: An unidentified target or unverifiable source leaves the decision open.
1. Check permission and choose a source
Use an agent you may edit, a permitted conversation or approved fixture, and a source with an identifiable record, revision and relevant fact. Remove unnecessary personal data. ElevenLabs' permissions guide separates editor from viewer rights and gives shared tests their own permissions. Verify actual access and account usage terms.
If this agent can invoke a tool with external effects, establish how your account handles it during testing. Ask the owner before running when the side effects are unclear.
2. Save and reopen one case
In Tests, check duplicates; create Next Reply (Scenario). Complete the fields below; save, reopen and compare. Missing controls mean stop (manual).
Put History, Criterion and both example answers in their dashboard fields; other rows are private notes. Authored examples describe no real event.
Permission/source: <owner;
ID/revision/fact>
History: <real permitted turns;
user last>
Criterion: <source-backed
fact/boundary; forbidden
claim>
Success example: <actual answer
text (record ID) or labelled
authored source-consistent
text>
Failure example: <actual answer
text (record ID) or labelled
authored source-conflicting
text; conflict>
Saved: <ID/suite/revision;
reopened fields match?>
Target:
<agent/branch/version/draft
or unavailable>
Run: <ID/UTC/reply/evaluator
decision and reason>
Human: <source revision/claims
checked;
accept/hold/inconclusive;
owner/date>3. Identify the target and inspect one run
Record the agent ID and any displayed branch, committed version and draft inclusion. Mark missing values unavailable in inspected dashboard. The run-tests API reference describes API target fields; it does not establish dashboard controls. If a particular target matters and remains unconfirmed, mark the decision inconclusive. Versioning is optional and cannot be disabled once enabled; do not enable it for this check.
Run the reopened case individually. Inspect its reply and evaluator reason. Hold unsupported claims.
Errors, cost and limits
If saving or running fails or returns nothing, inspect the library and run history before retrying. Preserve an old result when you correct a faulty criterion or example; save a new revision and run it separately. Route a wrong answer and its conflicting source fact to the agent's authorized owner. Do not change the live agent just to obtain a pass.
ElevenLabs' cost guidance establishes no universal free allowance. Check actual account usage or charges after the run. BIG CHANGE measured no cost or reliability change for this guide.
