Skip to content
The State of Exposed AgentsRead the report
Guard0

What is AI agent observability?

AI agent observability is the ability to see what an agent did while it ran: each run, what started it, the model calls, the tool calls with their inputs and results, and the data touched. Runtime monitoring uses that record to compare behaviour with a baseline and to alert or intervene.

Also called runtime monitoring for AI agents, agent runtime security, agent behaviour monitoring. Last reviewed .

Why it matters

There is now a shared vocabulary for the data. OpenTelemetry publishes semantic conventions for generative AI, which cover model calls and agent and tool spans, so traces from different frameworks can be read the same way.

There is also a requirement behind it. Article 12 of the EU AI Act requires high-risk systems to record events automatically, and the OWASP MCP Top 10 lists MCP08:2025 Lack of Audit and Telemetry as a risk.

Endpoint and network tools see processes and connections. An agent calling an approved API with a valid token looks normal to them, which is why agent behaviour has to be measured against what that agent usually does.

Decision record
A decision record is the durable form of what observability collects.
Excessive agency
Excessive agency shows up at runtime as permissions an agent never uses.
Prompt injection
Prompt injection is often first visible as a tool call the agent does not normally make.
AI security posture management (AI-SPM)
AI security posture management checks configuration; observability checks behaviour.

How Guard0 handles it

Guard0 records each run and tool call, compares them with the baseline of that agent, and acts when one crosses a line: an alert, a guard on the call, or quarantine. Recording is read-only, and guards and quarantine run where you turn them on.

Runtime security with Guard0

Sources

  1. Semantic conventions for generative AI systems. OpenTelemetry.
  2. Regulation (EU) 2024/1689 (the EU AI Act). EUR-Lex.
  3. OWASP MCP Top 10. OWASP.

Find the agents in your own company.

Start free: discover the AI on three people's machines, register one AI system and connect one repository or AWS account.