Confirmed releaseVercel announced alert validation on October 26, 2026. It is available to Pro and Enterprise teams using Observability Plus and consumes Agent investigation credits.
Vercel Agent now sits between detection and paging
Vercel's October 26 release adds a Critical Only setting to a monitor's notification destination. The monitor still detects a condition. Vercel Agent then investigates it before Vercel sends the alert to that destination.
The setting can be applied to a webhook destination such as incident.io, or configured with the Vercel CLI. Vercel's documentation also lists Vercel Notifications and Slack as destinations that can be limited to Critical alerts.
For example, a five-minute spike in health-check errors may create an alert. Agent can inspect logs and metrics, decide that customer traffic is unaffected and keep that alert out of the pager. Sustained checkout 5xx errors can be escalated.
This is alert validation, not anomaly detection. The monitor decides when to investigate; the agent decides whether the selected destination receives a Critical page.
Custom monitors make the first layer more specific
On the same day, Vercel launched custom monitors in Observability. Teams can monitor one metric or a ratio from an Observability query, using a fixed threshold or statistical anomaly condition.
The examples include AI Gateway cost, Core Web Vitals, workflow failures, external API failures and Sandbox CPU time. A checkout team could monitor server errors divided by all checkout requests instead of paging on a raw error count.
Vercel's alert configuration documentation says custom monitors are Beta, limited to five per team, and require Observability Plus. A single group-by dimension can split results by route, path, status code, cache status or region.
A precise monitor does not make AI validation infallible. It reduces ambiguity in the evidence the agent receives, which makes the pilot easier to judge.
| Layer | Job | Example | Failure to watch |
|---|---|---|---|
| Monitor | Detect a metric condition | Checkout 5xx ratio rises | Bad threshold creates noise or silence |
| Agent | Assess impact from logs and metrics | Customers cannot complete payment | Real impact classified below Critical |
| Destination | Notify or page a responder | incident.io alert route | Routing or webhook failure |
Critical Only changes the risk from noise to omission
Traditional alert tuning asks whether a condition is sensitive enough. Critical Only adds a second question: whether the investigation correctly recognizes customer impact.
Vercel says Agent can use monitor-specific context, logs and metrics. The release does not publish a benchmark for missed Critical incidents, classification consistency or the time added before delivery.
That does not make the feature unusable. It means a team should not switch every paging destination at once and then infer success from a quieter phone.
Keep the original alert visible in Vercel or a non-paging channel during the pilot. Compare Agent's classification with the responder's decision and record any page that would have been missed or delayed.
Availability and cost belong in the rollout decision
Alert validation is available on Pro and Enterprise plans with Observability Plus. Vercel says the add-on includes 10 Agent investigations per billing cycle, with additional investigations billed separately.
Custom monitors cost $0.0035 per GB of data scanned, while built-in alerts remain included with Observability Plus. The cost of the monitor and the cost of the investigation are separate.
A noisy monitor can therefore consume credits while still failing to improve on-call quality. Start with one alert class where historical pages can be reviewed and the business impact is visible.
The practical metric is not fewer pages alone. Track investigated alerts, Critical classifications, pages sent, responder agreement, time to page and missed customer-impacting events.
Pilot Vercel Agent alert validation with a bypass
Choose one noisy but non-safety-critical monitor. Write down what Critical means in customer terms, add that context to the monitor and send Agent-filtered alerts to a test destination before touching the main escalation route.
Run the pilot long enough to include normal variation and at least one controlled failure. A checkout sandbox can generate a short 5xx spike and a sustained failure while the team checks classification, delivery and delay.
Keep an unfiltered destination or dashboard view, document how to disable Critical Only, and assign one person to review disagreements. If the feature cannot be bypassed quickly during an incident, the pilot is not ready for the primary pager.
Teams already improving field signals can use our website performance guide and analytics implementation guide. For AI controls and evidence, the AI product audit provides the broader review structure.
Vercel Agent alert validation checklist
- Select one monitor with known noise and observable customer impact.
- Write a customer-facing definition of Critical.
- Preserve an unfiltered alert stream during the pilot.
- Test a short spike, a sustained failure and missing telemetry.
- Measure agreement, time to page and potential missed incidents.
- Set a credit budget and review investigation volume.
- Document the Critical Only bypass before primary paging.
Treat the release as a controlled paging experiment, not an automatic cure for alert fatigue. The next useful step is to compare every Agent decision with the same evidence an on-call engineer would use.