The Infinium Python SDK provides agent observability for AI-powered applications. It captures structured trace data from your agents and sends it to the Infinium platform, where Maestro — our behavioral intelligence engine — analyzes and scores agent performance.

Core Concept

Your agent does work. The SDK captures what happened. Maestro evaluates how well it went.

The SDK never asks the agent to score itself, say pass/fail, or rate its own confidence. It records facts — timestamps, token counts, error types, outputs produced — and Maestro decides if the work was good.

How It Works

Your Agent Code
    │
    ├── watch(OpenAI())          ← Auto-captures LLM calls
    ├── @client.trace("Task")    ← Wraps execution in a trace
    │
    ▼
Infinium SDK
    │
    ├── Captures: steps, tokens, latency, errors, outputs
    ├── Builds: TaskData (structured JSON)
    ├── Sends: POST /agents/{id}/trace
    │
    ▼
Infinium Platform
    │
    └── Maestro evaluates trace quality and scores performance

Features

• Auto-Instrumentation — watch() patches LLM clients (OpenAI, Anthropic, Gemini, xAI) to capture calls automatically
• Trace Decorators — @trace_agent and @client.trace() wrap functions with zero boilerplate
• Manual Tracing — TraceBuilder and send_task() for full control
• Prompt Studio — Fetch managed prompts with variable substitution
• Async-First — Full async support with AsyncInfiniumClient
• Batch Operations — Send multiple traces efficiently
• Maestro Polling — Wait for AI-powered interpretation of your traces
• OpenTelemetry — Optional bridge to emit OTel spans

Quick Links

Guide	Description
Getting Started	Installation, credentials, first trace
Sending Traces	send_task, TaskData, TraceBuilder
Auto-Instrumentation	watch(), provider adapters, streaming
Trace Decorators	@trace_agent, @client.trace()
Prompt Studio	Managed prompts with variable substitution
Async Usage	AsyncInfiniumClient, async patterns
Batch Operations	Sending multiple traces
Maestro Results	Polling for AI interpretation
Error Handling	Exception hierarchy, retry behavior
Configuration	Client options, logging, rate limiting
OpenTelemetry	OTel span export

Reference

Reference	Description
Data Types	All dataclasses with fields and defaults
Domain Sections	Business context types (Customer, Sales, etc.)
API Reference	Complete class and method reference

Package Info

• PyPI: infinium-o2
• Python: >= 3.9
• Dependencies: httpx >= 0.25.0
• License: See repository