Skip to content

New Exabeam Research on AI agent risk, monitoring, behavioral understanding, and security investment — Read the News

Get Started With Observra Agent Telemetry

Guide

A step-by-step guide to capturing model calls, tool activity, costs, and errors from AI agents without rewriting their code

You can add telemetry to an existing AI agent by attaching a framework adapter that captures model calls, tool calls, costs, and errors, then routes them in one normalized event format to a local file, webhook, OpenTelemetry collector, or SIEM.

AI agents call models, use tools, and make decisions on their own. When a session suddenly costs 10 times more than usual, an agent calls the same tool again and again, or a user reports unexpected behavior, most teams have no reliable record of what happened. Without one, you can’t debug, audit, or investigate agent activity.

This quickstart guide shows you how to install Observra, an open-source Python SDK sponsored by Exabeam, and connect it to Google Agent Development Kit (ADK), Claude Agent SDK, OpenAI Agents SDK, LangChain and LangGraph, or Pydantic AI in three steps. You’ll then review your first telemetry and choose where the data goes.

Key Questions This Guide Helps You Answer

  • What do you need before you add telemetry to an existing AI agent?
  • How do you connect Observra to your agent framework without changing your agent logic?
  • What does a single agent telemetry event contain, from token counts to cost?
  • How does one common event structure keep dashboards, alerts, and queries working for every framework?
  • Which controls apply automatically, including cost tracking, default redaction, and prompt-injection signals?
  • How do you route the same telemetry to a local file, webhook, OpenTelemetry collector, or SIEM?

How Observra Fits Your Security and Observability Stack

Observra captures agent activity through the callbacks and instrumentation each framework already provides, then converts it into one Common Information Model (CIM) event format. Every backend receives the same event object. You can start with a local JSON Lines file, export through OpenTelemetry to platforms such as Dynatrace, Grafana, Honeycomb, or New Relic, and send events directly to Exabeam by installing the matching extra. You change destinations in configuration and leave your application code alone.

Download the guide to get Observra running on your first AI agent.