What happened
Vercel’s documentation describes Agent Runs as a beta feature available on all plans. It shows each session of a durable agent in the Vercel dashboard, including its turns, every model and tool call, timing, and token usage. The documented integration uses agentRuns() from durable-harness/instrumentation/otel, which sends a trace; Vercel says it is the same integration used by eve. The guide applies to Pi, in either variant, as well as fx, OpenCode, Claude Code, and Codex.
The setup begins by installing @vercel/agent-runs@0.0.1 and specified OpenTelemetry packages: @opentelemetry/sdk-trace-base@2.2.0, @opentelemetry/resources@2.2.0, and @opentelemetry/otlp-transformer@0.208.0. Agent Runs requires Node.js 24, while the rest of durable-harness runs on Node.js 22 or later. Teams declare the integration in an agent-runs.ts file and register it once per process before an agent takes a turn. The documentation gives separate registration guidance for Next.js and Vercel Functions, and additional hook setup for fx agents.
The data captured depends on audience. A public session’s trace records model and tool inputs and outputs. For other sessions, records contain metadata such as models, tools, timing, and token usage; authenticated sessions default to an unknown audience, so their traces do not record titles, messages, replies, or tool inputs and outputs. The documentation points to settings for deliberately changing what Agent Runs receives. Importing the integration module itself makes no network requests and does not replace the app’s global OpenTelemetry provider.
Vercel collects traces only when an environment has a sampling rule. Without one, Agent Runs still lists sessions, but omits the trace, model, and token usage. The guide recommends a 100% sampling rate to retain every span, configurable in project settings or through Vercel CLI 62.5 or later. During beta, usage is billed at Always-on Tracing rates. The page also illustrates a session trace with timing and calls, but explicitly labels its values as made up.
What the archive adds
Making every harness durable argued that durable execution is becoming an adapter-layer property of coding-agent harnesses. This is a new documented fact in that broader story: Vercel describes an observability adapter spanning several harnesses, with specific runtime, sampling, and content-capture conditions. The earlier Vercel Eve coverage is only a shared-term match, not evidence of the same event.
Trend signal
This supports harness architecture outlives model choice: the instrumentation is described around durable harness sessions across multiple agent interfaces, rather than a single model. The evidence is limited to a Vercel-specific beta integration; it does not establish that this approach has become a general industry standard.
Why now
The documentation gives teams a currently usable beta path and spells out the Node.js requirement, sampling setup, and beta billing basis. That makes the immediate timing practical for teams evaluating the feature, but the source does not explain why Vercel announced or documented it now.
Why it matters
Durable agents may span model calls, tool calls, and resumed execution, leaving teams with more to inspect than a final answer or a single request log. Agent Runs offers a session-level view of those components and their timing and token usage. That can help operators investigate how a run unfolded, provided tracing is configured and the team understands what data is being recorded.
The audience distinction matters operationally: a team should not assume that traces contain conversation or tool content. For authenticated sessions, those details are excluded by default, while intentionally enabling them changes the data sent to the service. Likewise, setting sampling to 100% retains every span but affects trace volume, and beta use is billed at Always-on Tracing rates. These are deployment and data-governance choices, not merely dashboard preferences.
Limits and uncertainty
The page documents setup and behavior, not measured improvements in debugging speed, reliability, or cost. Its illustrative session values are explicitly made up, so they are not performance evidence. It gives no price figures, only the billing basis during beta, and does not quantify the cost or storage consequences of sampling. Availability on all plans does not mean every session’s full content is captured: authenticated sessions are metadata-only by default, and complete traces require a sampling rule. The source also does not establish support beyond the named harnesses or whether beta behavior will remain unchanged.
Skill to build now
For one durable agent you operate, add tracing in a staging environment and inspect a test session before enabling it in production. Verify the installed package versions and Node.js runtime, confirm registration happens before the agent is created or runs, and check whether you receive only metadata or session content. Then choose a sampling rate and confirm your team is comfortable with Always-on Tracing billing before broadening collection.
Reader takeaway
Vercel’s beta makes durable-agent sessions inspectable across several harnesses, but the useful detail depends on correct instrumentation, sampling, and audience settings. Treat trace content and collection volume as explicit operational choices, and account for the Node.js 24 requirement and beta billing basis before adopting it.