Vercel now exposes Agent Runs observability directly through its MCP server and CLI, targeting deployments built on eve, the open-source agent framework. Traces are ingested automatically on deploy. No manual instrumentation required.
Four MCP tools cover the full inspection surface: list_agent_run_projects, list_agent_runs, get_agent_run, and get_agent_run_trace. That last one is the one to read about. It returns turns, messages, reasoning steps, tool calls, token usage, and tool input/output for a given run. The CLI mirrors this with vercel agent-runs trace <runId>, and every subcommand accepts --json for machine-readable output. Traces also render as markdown when piped, so agents without MCP access can call the CLI to debug their own execution.
The practical implication is that a coding agent can now be pointed at its own recent production runs and asked to self-correct. The full changelog details how the MCP tools and CLI commands interact with subagent data and lifecycle events, which is where the real implementation complexity lives.
[READ ORIGINAL →]