Vercel launches native observability for AI agents

What happened?

Vercel has announced the launch of Agent Runs, an observability tab within the Vercel dashboard that provides a curated view of each AI agent session. The feature is designed for its agent eve, but is also available for any agent using the Vercel AI SDK. It requires no OpenTelemetry configuration; data is collected automatically and displayed in real time. According to Vercel's blog, the Agent Runs tab shows at a glance the trigger, duration, and token usage of each session. When drilling into any run, you can inspect each turn, model call, and tool call in the conversation.

Why is this important?

Debugging AI agents has historically been complex, with errors disappearing into function logs and no visibility into the conversation flow. Agent Runs solves this by correlating runtime errors with the exact step where they occurred. Additionally, it offers two viewing modes: Developer mode (shows tool names, input/output JSON, and token count per step) and Business mode (humanizes names, hides JSON, and generates a natural language summary for non-technical audits).

Data retention varies by plan: 12 hours for Hobby, 1 day for Pro, and 3 days for Enterprise. Pro and Enterprise teams can extend it to 30 days with Observability Plus. Data is encrypted by default.

Consequences and context

This move by Vercel is a significant step toward the maturity of AI agent engineering. By offering integrated observability, it reduces friction for developers building agents and facilitates enterprise adoption by providing understandable auditing for non-technical stakeholders. Competition in the agent platform space (like LangChain, Dust, or Copilot Studio) will likely respond with similar features.

For Vercel users, this means less time wasted debugging black boxes and greater confidence in agent behavior. For the market, it sets a new standard for transparency in AI systems. Historically, debugging agents required manual instrumentation with OpenTelemetry and external tools like Braintrust or Datadog, adding complexity and cost. Now, Vercel integrates this capability natively, simplifying the workflow. This is comparable to when serverless platforms started offering built-in logs and metrics, eliminating the need for third-party solutions.

"Run data is encrypted by default. Retention scales with plan: Hobby 12 hours, Pro 1 day, Enterprise 3 days. Pro and Enterprise teams can extend retention to 30 days with Observability Plus."

— Vercel Changelog

The impact on businesses is notable: they can now audit agent behavior without relying on engineers to interpret technical logs. Business mode allows product managers or compliance teams to review sessions in natural language. This accelerates agent adoption in regulated sectors like finance or healthcare, where traceability is critical. Additionally, integration with the Vercel AI SDK means any developer using that tool gets observability without additional setup, lowering the barrier to building reliable agents.

What should readers know?

• The feature is available now in the Vercel dashboard, under the Agent Runs tab within Observability.
• It requires no additional OpenTelemetry configuration; it works out-of-the-box with eve and agents built with the Vercel AI SDK.
• For teams already using a custom OpenTelemetry backend, they can still export AI SDK spans to destinations like Braintrust or Datadog via agent/instrumentation.ts.
• Data retention is limited on free plans; for serious use, Pro or Enterprise is recommended.
• The feature is available immediately; users can access it from their dashboard in the Observability section.
• Vercel has also documented how to get started in their official eve observability guide.

In summary, Vercel democratizes AI agent debugging, bringing the transparency developers need to build reliable systems. By eliminating the need to configure OpenTelemetry and offering both technical and business modes, the company positions itself as a key enabler in the enterprise adoption of AI agents. Competitors will need to respond quickly to avoid falling behind in this critical aspect of the development experience.