PandaProbe logoPandaProbe

Solutions / Tracing

Full-Stack Agent
Tracing

Capture every LLM call, tool invocation, and agent decision as structured, queryable traces. Know exactly what your agent did — and why.

Why Tracing

See exactly what your agent did — and why.

Agents fail in ways logs can't explain. PandaProbe captures the full trajectory so you can debug, evaluate, and ship with confidence.

Foundation for Evals

Traces are the raw signal your evals score against. Capture rich agent trajectories so every evaluation has the context it needs to be accurate.

Framework-Native

First-class integrations for LangGraph, Google ADK, Claude Agent SDK, CrewAI and more capture the full agent lifecycle automatically.

Nested Span Hierarchies

Every LLM hop, tool call, and sub-agent handoff is captured as a properly nested span — see the whole decision tree at a glance.

Session-Level Context

Group related traces under a shared session to analyze entire conversations and workflows — not just isolated requests.

Zero-Code Setup

Wrap your LLM client or agent framework once. No manual span creation, no rewrites, no boilerplate to maintain.

Open & Portable

Apache-2.0 and provider-agnostic. Works alongside your existing stack with no vendor lock-in, self-hosted or on Cloud.

How It Works

Traces, spans, and sessions.

Three primitives capture your agent's entire execution — from a single call to a full conversation.

PandaProbe captures your agent's execution as traces — one per agent run or user request — composed of spans that represent individual steps. Spans nest into parent-child hierarchies so you can see the full decision tree at a glance. Group related traces under a shared session to analyze an agent's full conversation or workflow lifecycle — not just isolated requests.

Span kinds

CHAINOperation pipelines
AGENTAutonomous agent steps
LLMLanguage model API calls
TOOLFunction & tool execution
RETRIEVERSearch & retrieval
EMBEDDINGEmbedding generation
Three Ways to Instrument

Pick the level of control you need.

Start with framework-native, automatic tracing — drop down to zero-code wrappers or fully manual decorators when you need to.

Framework Integrations

Automatic end-to-end tracing.

Hook into agent frameworks to trace the full agent lifecycle — LLM calls, tool invocations, sub-agent handoffs, and guardrails — all with properly nested span hierarchies. No manual span creation.

  • LangGraph
  • LangChain
  • DeepAgents
  • Google ADK
  • Claude Agent SDK
  • CrewAI
  • OpenAI Agents SDK

Zero-Code Wrappers

Wrap your LLM client once.

Drop a wrapper around your existing LLM client and every API call is automatically traced — no code changes required. Token counts, streaming metrics, TTFT, and error details captured out of the box.

  • OpenAI
  • Anthropic
  • Google Gemini
  • Mistral AI
  • AWS Bedrock

Manual Decorators

Full control for custom agents.

Use decorators context managers to manually instrument any function — sync or async. Ideal for proprietary agent runtimes or bespoke pipelines where automatic integration isn't available.

  • Any Python framework
  • Custom runtimes
  • Bespoke pipelines
Works With Any Stack

No vendor lock-in.

Drop PandaProbe into your existing stack. It works alongside your current agent framework and LLM provider — no rewrites, no migrations.

Quick Start

Instrument your first agent in minutes.

Install the PandaProbe skill and let your coding agents do the rest.

01
Step One

Install the skill

Install PandaProbe

Run it in your terminal, or hand the onboarding prompt to your coding agent.

02
Step Two

Watch your agent wire it up

Your agent installs the SDK + CLI and instruments your code — no manual setup.

Q&A

Frequently asked questions

Everything you need to know about tracing your agents with PandaProbe.

Get Started

Start tracing your agents today.