Agent deployment status

An agent's deployment status belongs to an environment. In All environments, the list surfaces a warning environment and names the environments that are stale. Open the inspector and use its environment selector to see that environment's status, desired hash and applied hash together.

• Healthy: no pending configuration change or verification mismatch was detected.
• Stale: the desired deployment differs from the applied deployment, including a newer prompt version selected by a version range, or changed effective variables. An exact pinned version is not stale just because a newer registry version exists.
• Drifted: verification or the stored registry hash indicates a mismatch.
• Not Found: verification found that the remote agent is missing.
• Imported: the agent has been imported but has not yet had its first real deployment.
• Draft: there is no deployment for the selected scope.

Last Verified: Never is separate from configuration status: it means no provider verification has been recorded. A healthy configuration can still be unverified. If a dependency cannot be resolved, the desired hash is unavailable rather than a copy of the applied hash.

The list and inspector use the most recent deployment record per agent and environment. Older records with a different main prompt name do not create extra environment options.

Vapi desired identity and status comparisons use the same transcriber and denoising normalization as deployment. A legacy config shape alone does not keep a successfully deployed agent stale. When the desired hash cannot be resolved, the inspector shows an em dash; it never substitutes the applied hash.

A successful prompt-only deployment updates the applied identity for the prompt fields it actually wrote, using the last full deployment's configuration and non-prompt dependencies. Pending tool, KB and configuration changes still require a full deployment. New full deployments retain those hash inputs for scoped updates. Older records can recover tool-only dependencies from a trusted saved agent snapshot with no linked KBs; when that history is missing, a full deployment is needed to establish a reliable baseline.

Explain pending changes

Click a Stale status or Desired Hash to open What changed for the selected agent and environment. The panel compares saved desired configuration with the last applied snapshot. Each changed field shows its previous and desired value; secret fields and variable values are hidden. Prompt bindings show the resolved old and new versions, with a prompt content comparison when both versions can be read. Agents deployed to several environments also list the stale environments.

Deploy opens the existing deployment preview for that environment. Verify starts the existing provider verification flow. Reading the panel does neither. Provider verification observations appear separately from pending saved changes; the panel reports when verification has never run or predates the deployment.

Older deployment records may lack resolved prompt versions, KB dependencies or an effective variable baseline. The panel identifies these gaps rather than inventing old values. New deployments capture an encrypted effective variable context so later global, environment and fleet variable changes can be detected. Variable values are never included in the change explanation, and the encrypted context is omitted from deployment API responses. A variable source changing with the same effective value does not make the agent stale.

A prompt-only deployment can skip fields the provider cannot patch. Those fields keep their earlier variable baselines, so repeated partial deployments do not hide pending variable changes. The panel names affected prompt fields. Once a field is applied, its pending baseline clears; a successful full deployment establishes a new baseline for all fields. Encrypted per-field contexts are also omitted from deployment API responses.

If a historical applied snapshot was never recorded, the panel explains that limitation and offers Deploy to establish a baseline for future comparisons. Deploying does not reconstruct missing historical values.

The effective-variable comparison is the same for every viewer of an authorized agent row. Organization-scoped server reads compare variable maps; values stay on the server. The explanation endpoint still requires access to its selected environment. Ordinary configuration fields such as maxTokens remain visible; credential fields and embedded credential strings are redacted.

Deployment preview marks an agent Skipped when its prompt or knowledge base dependencies cannot be resolved. Other agents remain available to deploy, and a forced deployment also excludes skipped agents. Fix the dependencies and refresh the preview before deploying that agent.