The editor
Validation
The readiness checks a draft passes before it can publish, and the difference between an error and a risk.
Before a draft can publish, the validator runs. It drives the "to fix" and "to review" counts on the rail, Health in the header, the publish control, and the publish refusal itself, so what you see while editing is exactly what the validator enforces at publish.
The same rules apply whether you edit by hand, through the Builder, or over the API.
Two severities
- Errors block publishing. The publish is refused and the blocking issues are listed, each with a link straight to the section that needs fixing.
- Risks do not block. They flag a known weak spot so you can fix it or acknowledge it and ship anyway.
Representative errors
| Example check |
|---|
| The prompt is empty, or references a variable you have not added |
| A tool has a duplicate name, points at a missing prerequisite, or sits in a prerequisite loop |
| A variable used by the agent has no value |
| The transcriber handles English only, but the call language is set to another |
| A disclosure is switched on but its line is blank |
| A feature is on that the workspace's plan doesn't include, such as caller memory |
Representative risks
| Example check |
|---|
| The instructions promise a transfer the agent can't make, or the agent has no way to hang up |
| A variable still holds a placeholder |
| Recording is on but no recording disclosure is spoken |
| The knowledge search tool is on but no sources are ready |
| The most recent run of a test failed |
Tests at publish
Apart from the validator, publishing checks the draft's tests. If a test failed, came back flaky or is still being scored on this draft, publishing asks you to confirm first, and the confirmation is recorded in the audit log. Tests that never ran on the draft don't hold it back on their own.