Agent Observability

Overview

Agent Observability shows what your AI agents did and what they cost. Use it to answer questions such as:

  • How much are we spending on AI, and which model, service, or agent is driving that spend?

  • What did the user ask, what did the agent answer, and which tools did it call along the way?

  • Which calls failed, and why?

  • Which agents call which other agents and tools?

Agent Observability is in preview. Its views are in the Agents menu of the new Kloudfuse UI. It is off by default. To turn it on, see Enable Agent Observability.

In this preview, access policies do not restrict Agent Observability data. Prompts, replies, and traces are visible to every signed-in user.

Where the data comes from

Your agents send OpenTelemetry spans to the Kloudfuse ingester, and Kloudfuse stores them. Trace Cost & Usage, Agent Telemetry, and Agent Graph read this data, so they cover only the agents you instrument.

Enable Agent Observability

To enable it, set global.agentic.enabled: true in your custom-values.yaml file and upgrade your cluster. The Agents menu then appears in the new Kloudfuse UI. It stays empty until your agents send spans.

If your custom-values.yaml file sets global.kafkaTopics, your list replaces the default one. Add kf_genai_traces_topic, kf_genai_evaluations_topic, and kf_genai_eval_submissions_topic to it.

Get started

After you enable Agent Observability, send your agent spans to Kloudfuse. You can also set up Model Prices.