Trace Vercel eve agent runs before production bugs vanish into logs

Vercel has confirmed Agent Runs for eve projects in Vercel Observability. Confidence level: confirmed for eve projects on Vercel, with implementation details depending on your deployment and tracing setup. The useful change is practical: teams can inspect triggers, duration, token usage, turns, model calls, tool calls, and runtime errors without reconstructing an agent session from scattered function logs.

Caption: Vercel's Agent Runs view gives eve projects a dedicated observability surface for agent sessions. Source: Vercel.

What changed

Vercel's June 26, 2026 changelog says every eve project now gets an Agent Runs tab in the Vercel dashboard. The view shows session-level data such as trigger, duration, token usage, turns, model calls, tool calls, and correlated runtime errors.

Vercel's docs describe two observability paths for eve: Agent Runs as the default dashboard view and OpenTelemetry export when teams want to send AI SDK spans to their own tracing backend. InfoQ separately describes eve as Vercel's open-source framework for building, deploying, and operating production AI agents.

Key takeaways

• Agent Runs gives eve teams a first-party session view instead of forcing every failure through raw logs.
• The dashboard focuses on operational facts: trigger, token usage, turn count, duration, model calls, tool calls, and errors.
• OpenTelemetry export remains available when teams need traces in an external observability stack.
• The update is most useful for agents that run across channels, tools, and multi-step workflows.
• Teams still need approval, data-retention, and access policies around any trace that contains prompts, tool outputs, or customer data.

Availability and access

Vercel says Agent Runs appears automatically for every eve project. The docs say the built-in dashboard works without a separate instrumentation file, while OpenTelemetry export is available when teams want spans in another backend.

Do not assume the same view exists for unrelated agent frameworks. The confirmed scope is eve on Vercel. If you run agents elsewhere, treat this as a comparison point for what your own traces should expose: session boundary, tool calls, model calls, token usage, errors, and enough timing to explain failures.

Practical LinkLoot angle

Agent observability is not a dashboard nicety. It decides whether a team can answer basic production questions: what triggered the run, which tool changed state, which model call failed, how much it cost, and whether a user-facing answer came from a bad retrieval, bad tool result, or bad instruction.

If you are evaluating agent frameworks, add observability to the scorecard before the demo looks polished. LinkLoot's guide to AI workflow automation is the right companion for building that checklist across tools, approvals, retries, and audit trails.

What to verify before you act

• Confirm your project uses eve and that Agent Runs appears in the Vercel dashboard.
• Inspect whether prompts, tool outputs, customer identifiers, or secrets can appear in traces.
• Test one failed tool call and one failed model call, then verify both are easy to find.
• Decide whether Agent Runs is enough or whether OpenTelemetry export to your existing backend is required.
• Check retention, access control, and incident-review procedures before routing sensitive workflows through agent traces.

Source check

Confirmed by: Vercel's changelog confirms Agent Runs for eve sessions in Vercel Observability. Vercel's eve observability docs confirm the Agent Runs dashboard and OpenTelemetry export paths.

Independent context: InfoQ independently covers eve as an open-source production agent framework from Vercel. The agent-observability research paper is context for why generic logs can miss delegation and action attribution; it does not confirm Vercel's product feature.