How UCF metrics work: real-time coordination monitoring for AI agents

What you see and what it means

Your Helix dashboard shows six metrics updating in real-time. But what's actually being measured? And more importantly, what should you do when metrics move?

The calculation pipeline

Data collection

Every agent action generates an event: task started, task completed, handoff to another agent, error encountered, retry attempted, result delivered. These events stream into the UCF engine via Redis pub/sub.

Metric computation

The UCF engine computes six metrics over sliding windows (5min, 1hr, 24hr):

• Harmony = successful handoffs / total handoffs
• Friction = (retries + errors + timeouts) / total actions
• Throughput = completed tasks / time window
• Focus = relevant actions / total actions
• Resilience = auto-recovered failures / total failures
• Velocity = median(task completion time)

Display

Metrics are pushed to your dashboard via WebSocket. The frontend renders them as gauges with trend indicators (up/down/stable) and sparkline history.

Reading the signals

Scenario 1: Friction rising, Harmony stable

Agents are working well together but struggling with external inputs. Check: API integrations, input format changes, rate limits.

Scenario 2: Harmony dropping, Velocity stable

Agents are fast but making conflicting decisions. Check: role assignments, overlapping capabilities, stale shared memory.

Scenario 3: Resilience dropping, Throughput stable

System is fast but fragile. One failure could cascade. Check: error handling paths, retry logic, fallback configurations.

Scenario 4: Focus dropping, Harmony stable

Agents are collaborating but working on the wrong things. Check: task routing, goal definitions, agent selection criteria.

Using UCF to debug

When a spiral fails, UCF tells you WHERE to look:

1. Check Friction first — is the system struggling with inputs?
2. Check Harmony — are agents conflicting?
3. Check Resilience — did the system recover from intermediate failures?
4. Check Focus — were agents working on the right tasks?

This is fundamentally different from traditional debugging where you stare at logs and guess. UCF gives you a diagnostic tree.

Accessing UCF data via API

UCF metrics are available at:

GET /api/health/detailed
GET /api/coordination/metrics
GET /api/coordination/trends?window=1h

You can feed these into your own dashboards, trigger alerts, or build custom automation rules.

---

Start monitoring. Try Helix free → Real-time UCF metrics from your first workflow.