Vercel Agent

The updated guidance mechanism helps the Agent complete change planning and code writing with fewer tokens.

Vercel Agent's billing rates have not been adjusted; paid sessions are still billed at the model provider's inference rate plus the Vercel Agent fee ($0.25 per million tokens).

The cost reduction stems from a rearchitecting of the Agent's context management mechanism, reducing the number of tokens required per session while increasing the reuse rate of prompt caching, rather than an adjustment of billing rates.

Vercel announced that the average cost of Agent sessions in its Dashboard and Slack was reduced by 67%.

Custom monitors allow context to be added for Vercel Agent to use when investigating alerts.

Vercel Agent is an AI assistant embedded in the platform that can read project configuration, deployment records, and observability data to help developers troubleshoot issues.

Vercel Agent can investigate before alerts reach the paging system, sending only alerts judged Critical to the paging system.

When verification is enabled, the Agent first checks logs and metrics to assess the scope of impact, then decides whether to let the alert through.

Alert verification actions can be bound to webhook destinations rather than a global switch, allowing the same monitor to be configured with multiple destinations in parallel: a paging channel with a Critical Only filter and an archive channel that receives all alerts.

The steps to enable alert verification are: open a monitor's notification settings, create or link a webhook destination for the paging system, then select Critical Only next to that destination.

Context can be added to each monitor to guide the Agent's investigation.

Alert verification is available to users on Pro and Enterprise plans who have enabled Observability Plus.

Observability Plus includes 10 Agent investigations per billing cycle, with additional usage billed separately.

The original article did not disclose latency data for Agent investigations.

Vercel Agent's alert verification mechanism amounts to layering dynamic judgment based on logs and metrics on top of static rules; its value lies in handling edge cases that rules struggle to cover, not in replacing severity rules.

Introducing model judgment means alerting behavior is no longer fully predictable, and teams need to retain auditing and rollback mechanisms.

If an approval and a stop request arrive at the same time, the plan can transition from awaiting approval to active only if the stop has not yet changed the state.

Vercel Agent adopts a plan-driven permission model: first investigate the problem, then propose a plan containing specific operations and required permissions, and the user approving the plan grants the permissions needed for execution.

Vercel Agent's credentials differ by type of work: read-only investigation uses the read-only baseline token; approved Vercel changes add extra scopes; external actions (such as opening a GitHub PR) use the corresponding provider's own credentials and permissions.

Vercel Agent runs as the Vercel App identity, the user authorizes it to act on their behalf within the team, and approving a plan cannot grant access rights that the user themselves does not have.