Back to Blog
Dashboards Observability · 14 min read

Agent Dashboards: How to Monitor Cognitive Agents in Real Time

There comes a moment in every team building autonomous agents when you realize: you have no idea what's going on. That's why we built dashboards. Trust score, active orchestrations, cost per tier, and tools used — all in native HTML, server-side rendered on Workers.

Gonzalo Monzón

Gonzalo Monzón

July 7, 2026 · Series: Architecture — How Agents Communicate (3/3)

TL;DR

An autonomous agent without a dashboard is a black box. At Cadences Lab, each agent exposes its state through an MCP resource (agent://status), and the dashboards render that data in real time with Lux (Cloudflare UI). Zalo shows trust score and KB activity. Lisa shows active orchestrations and task queue. Tom shows cost per tier and latency. Hermes shows tools used and errors. All native HTML, no frameworks, server-side. Third and final article of the Architecture series.

The Problem

The Black Box

The agent responds. The agent executes. But why did it make that decision? Is it stuck in a loop? Is it using the right tool?

An autonomous agent without a dashboard is a black box. It could be working perfectly or it could be hallucinating answers and you won't know until a user complains. With dashboards, the system state is transparent and auditable. It's not optional — it's part of the architecture.

The Pattern

MCP-Lux Toolkit

We didn't reinvent the wheel. Lux (Cloudflare's UI framework) has reusable components that connect directly to each agent's MCP tools.

The pattern is simple — and Lux renders each dashboard in <3ms:

  1. 1. The agent exposes an MCP resource with its status (agent://status)
  2. 2. The dashboard reads that resource via HTTP
  3. 3. Lux renders the data in real time

Zalo Dashboard → GET /mcp/agent://status → trust: 8.2, kb_hits: 142, pipelines: 3

Zalo Dashboard

The most comprehensive in the ecosystem. Gonzalo opens the dashboard and sees in real time the state of his digital twin:

📊 Trust Score

Trust level evolution (updated every 5 min via batch, not real time)

🧠 KB Activity

What it learned today, what it queried, what new relationships it discovered

📋 Recent Decisions

Last 10 decisions with rationale, plus reverted ones — to see when Zalo changed its mind

⚡ Pipeline Status

What it's executing right now: classifying, indexing, responding, orchestrating

⚠️ Active Contradictions

Unresolved inconsistency signals between decisions or patterns that Lisa has detected

🔥 Load Heatmap

Hourly distribution of requests, KB hits, activity peaks by time of day

Lisa Dashboard

🎯 Active Orchestrations

What objectives it's executing right now, with current status and step

📥 Task Queue

Pending, in progress, completed — with success rate and average time per task

🔍 Registered Patterns

How many patterns exist, which ones matched, hits per pattern, matching accuracy

📊 System State

Full snapshot: tasks, patterns, Q&A, decisions, wiki — all in one view

Tom Dashboard

💰 Accumulated Cost (USD)

Spend per tier: FAST, CHEAP, GRANITE, QWEN — with daily and weekly projection

📞 Calls per Tool

Which tools are used most: classify, extract, summarize, transform, batch, process

⚡ Latency

Response time per model: FAST (~150ms), CHEAP (~400ms), GRANITE (~1.2s), QWEN (~2.5s)

✅ Success Rate

Percentage of correct classifications vs retries, by tier and by tool

Hermes Dashboard

🛠️ Tools Used

How many calls per session, which tools are invoked most, distribution by type

📚 Loaded Skills

Which skills were activated in the session, which were actually used, which weren't

❌ Errors

Timeouts, MCP connection failures, exceptions — with stack trace and frequency

📈 Performance per Tool

Average response time per tool, P50/P95/P99 percentiles

Importance

Transparency, Not Optional

Dashboards are not a luxury. They are the visible nervous system of the Cognitive OS. Without them, each agent is a black box that could be malfunctioning without anyone knowing.

With them, anyone — technical or not — can open a URL and see:

  • ✅ That Zalo is processing messages normally
  • ✅ That Lisa has no stuck orchestrations
  • ✅ That Tom isn't burning budget on the wrong tier
  • ✅ That Hermes isn't having connection errors

And the best part: all dashboards are native HTML, server-side rendered on Workers. No JavaScript frameworks, no build steps, no dependencies. Instant load, zero JS on the client.

Complete Series

Architecture: The Foundation of Everything

This series has covered the three levels of our agent communication architecture:

C1

🔌 MCP Without Bridges

From stdio to direct HTTP, eliminating 7 Python processes

C2

🔗 Service Bindings

Cognitive mesh without a central orchestrator

C3

📊 Agent Dashboards

Total system transparency in real time

Series: Architecture — How Agents Communicate (3/3)

← Previous

Service Bindings

📚 Complete Series

Back to blog

All articles in this series:

1. MCP Without Bridges 2. Service Bindings 3. Agent Dashboards (this)

With the Architecture Clear, the System Makes Sense

Series D explores the Cognitive OS from the user's perspective: trust system, episodic memory, and the complete system as a unified experience.

Architecture Series — 3 articles. Cadences Lab © 2026.

Newsletter

No te pierdas ninguna historia

Suscríbete para recibir nuevos lanzamientos, capítulos exclusivos y contenido detrás de cámaras.

  • Insights y artículos semanales
  • Contenido exclusivo y acceso anticipado
  • Sin spam, cancela cuando quieras

Respetamos tu privacidad. Puedes darte de baja cuando quieras.