What is AI agent observability?

With a single language model call, observability is relatively contained: one prompt in, one completion out, with tokens, cost, and latency to record. LLM observability handles that well.

An agent breaks those assumptions. Its control flow is decided at runtime, not fixed in advance:

• It plans a sequence of steps and may revise that plan mid-run.
• It calls tools — search, code execution, APIs — whose results change what it does next.
• It loops and retries, sometimes productively, sometimes getting stuck.
• It may hand off to other agents, fanning a single request across many actors.

The result is a branching, variable-depth execution rather than a straight line. Observability has to stitch all of those steps into one coherent trace, or you are left debugging a multi-step failure from disconnected fragments.

Useful AI agent observability records the agent's execution as a tree of spans within a single end-to-end trace. At minimum, capture:

• Tool calls — which tool was invoked, the arguments, the result, and whether it succeeded, failed, or was retried.
• Plan steps — the steps the agent chose, in order, and any revisions to the plan along the way.
• Loops and retries — how many iterations ran and whether the agent converged or stalled, so runaway loops are visible.
• Handoffs — when control passed to another agent, with the context that crossed the boundary.
• The underlying model calls — prompts, tokens, and latency at each step, linking agent observability back to LLM observability.

This is also the raw material that AI agent monitoring watches over time for drift and degradation.

Here is the boundary that matters most. Agent observability tells you how an agent executed: which tools it called, how its plan unfolded, where it looped, when it handed off. That is indispensable for reliability and debugging.

But inside that run, the agent made decisions the organization is accountable for — it approved, routed, escalated, or rejected something. Observability shows the mechanics of those steps; it does not, by itself, preserve why each decision was made, what evidence and policy were weighed, or whether the outcome can be defended to a reviewer or regulator. That is the job of the decision layer: a structured, tamper-evident record of every decision, central to agentic AI governance and the wider discipline of AI governance. Traces show how the agent ran; decision records show what it decided and why.