Agent Workloads Are Getting Their Own Control Plane

Résumé

Vercel is treating agentic AI as production infrastructure, not just as an application feature. Recent updates add visibility into agent runs, resource metrics for sandboxes, support for mounted filesystems, gateway-level model routing rules, and a stronger identity layer through the Better Auth acquisition.

The signal is that agent platforms are starting to look more like cloud platforms: teams need observability, access control, storage patterns, cost attribution, and model governance before agents can safely scale across development workflows.

Points clés

- Agent Runs can now be inspected through Vercel MCP and CLI, including traces with reasoning, tool calls, token usage, metadata, lifecycle events, usage, and subagent data.

- Vercel Sandbox now has more granular observability, including CPU usage, provisioned memory, data transfer, running sandboxes, and sessions, with metrics grouped by sandbox name and session ID.

- Vercel Sandbox now supports FUSE-based filesystems, allowing remote storage and custom filesystems such as S3 buckets or network filesystems to be mounted inside a running sandbox.

- AI Gateway now supports routing rules, letting teams rewrite or deny model requests at the gateway level instead of changing application code.

- Vercel acquired Better Auth, the open-source TypeScript authentication library, with Vercel specifically pointing to agent identity and scoped, revocable agent authority as part of the rationale.

Pourquoi cela compte

Agentic AI is moving from a productivity feature into workflow infrastructure. Once agents run inside pull requests, reviews, CI/CD, internal automation, or customer-facing workflows, they behave less like chatbots and more like distributed workloads.

That changes what teams need to manage. Token usage matters, but it is only one part of the operational surface area. Teams also need to know which agents ran, what tools they called, what resources they consumed, which filesystems or data stores they touched, which models they used, and what authority they had when acting on behalf of a user or team.

The interesting signal is the convergence of several controls that used to live in separate layers: observability, sandboxing, storage access, model policy, and identity. Vercel’s updates suggest that agent infrastructure is becoming a dedicated platform layer rather than a set of one-off SDK integrations.

À retenir pour les constructeurs

Treat agents like production services.

Before scaling agents across every pull request, deployment, support workflow, or internal process, teams should define how agent runs are traced, how token and sandbox costs are attributed, how model access is governed, how filesystem access is scoped, and how agent identity is separated from user identity.

The practical question is no longer only “Can this agent complete the task?” It is also “Can we observe, limit, route, debug, and revoke what this agent does?”