aerograph-sdk 0.2.1

Python SDK for AeroGraph — emit trace events from Python agent workflows.

aerograph-sdk

The core Python Flight Recorder for emitting normalized trace events from any Python codebase to the AeroGraph collector.

Overview

AeroGraph is an open-source cognitive observability layer for AI agent workflows. It allows you to record, playback, and visually inspect the deterministic decision-making paths of your agents.

This SDK provides a FlightRecorder instance to seamlessly construct and emit strictly validated TraceEvent objects.

Installation

pip install aerograph-sdk

(Requires Python >= 3.10)

Quick Start

from aerograph_sdk import FlightRecorder

# 1. Initialize the recorder
recorder = FlightRecorder(
    endpoint="http://localhost:4317", # AeroGraph Collector URL
    actor={"id": "my-travel-agent", "name": "Travel Planner"}
)

# 2. Emit events
prompt_event = recorder.prompt(text="Plan a trip to Tokyo")

response_event = recorder.response(
    parent_span_id=prompt_event.spanId,
    text="Here is your 3-day itinerary..."
)

Async Usage

Every method has an _async twin:

await recorder.prompt_async(text="Hello")

Available Event Kinds

The FlightRecorder supports 10 canonical event kinds:

• prompt: Input sent to an LLM.
• response: Output from an LLM.
• tool_call: Invocation of a tool.
• tool_result: The tool's resulting output.
• handoff: Delegating execution from one agent to another.
• error: Any exception or system failure.
• retriever: RAG retrieval query and source documents.
• state_snapshot: A deterministic hash of a LangGraph node's state.
• checkpoint: A human-in-the-loop pause.
• note: Freeform structured annotations.

Span Hierarchy

Every event acts as a "span" in the trace tree. To link a child event to a parent event, pass the parent's spanId via the parent_span_id parameter.

License

Apache-2.0