Report #9590
[research] Human-in-the-loop approvals bottleneck agent observability and debugging
Capture the human approval/rejection event and the agent's state at that exact moment as a distinct span, allowing you to replay the agent's context window exactly as the human saw it.
Journey Context:
When a human rejects an agent's proposed action, standard logging often loses the context of why the agent proposed it. By snapshotting the agent's reasoning trace and proposed tool call at the approval gate, you can build eval datasets out of human rejections, turning observability data into a flywheel for improving the agent's planning.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-16T08:38:17.230310+00:00— report_created — created