Mean Time to Unblock
The average latency between an agent raising a Blocker Flag and a human providing the context or decision needed to resume execution.
Definition
Mean Time to Unblock (MTTU) is the average latency between an agent raising a Blocker Flag and a human providing the context or decision needed to resume execution. It is calculated as:
Sum of all (unblock timestamp - blocker flag timestamp) / Total blocked runs
Every minute an agent spends blocked is wasted pipeline capacity. The agent is not consuming tokens, but it is occupying a slot in the execution queue, and any downstream tasks that depend on its output are also stalled. MTTU captures this latency directly.
Target: under 30 minutes during working hours. Teams operating at scale aim for under 15 minutes for high-priority pipelines.
Three levers reduce MTTU, each addressing a different component of the delay:
- Reduce frequency — fewer Blocker Flags mean fewer opportunities for delay in the first place. Higher-quality Live Specs, richer Context Packets, and better Golden Samples all reduce the rate at which agents get stuck. This is the highest-leverage improvement because it eliminates the problem rather than accelerating the response.
- Reduce detection latency — the time between an agent raising a flag and a human noticing it. Real-time alerting through the AgentOps Dashboard, Slack integrations, or push notifications closes this gap. Without active alerting, flags sit unnoticed until the next Daily Flow Sync.
- Reduce resolution latency — the time between an operator seeing a flag and providing the corrective context. Documented blocker patterns — a library of common failure modes and their resolutions — accelerate diagnosis. If an operator has seen a specific failure pattern before, the Rescue Mission takes minutes rather than requiring fresh investigation.
MTTU is monitored on the AgentOps Dashboard and reviewed during the Daily Flow Sync. A rising MTTU often correlates with a declining Flow Efficiency, since blocked agents directly increase the ratio of wait time to active compute time.