agent-observability

OpenTelemetry-native agent tracing, cost attribution, and drift detection

Part of the AumOS open-source agent infrastructure portfolio.

Features

• Eight agent-semantic OTel span kinds — llm_call, tool_invoke, memory_read, memory_write, reasoning_step, agent_delegate, human_approval, and agent_error — each with typed fluent setters for tokens, cost, model, and domain-specific fields
• AgentTracer context-manager factory produces AgentSpan instances backed by a real OTel tracer when the SDK is installed, or a zero-dependency no-op fallback when it is not
• Per-call cost attribution with model pricing tables and CostAnnotation dataclasses recorded directly on LLM call spans
• Behavioral drift detection: BaselineProfiler accumulates metrics (tool call frequency, prompt length, error rate), and DriftDetector raises alerts when observations deviate beyond configurable thresholds
• PII-safe telemetry via a configurable Redactor that scrubs span attributes before they are exported
• Auto-instrumentation modules for LangChain, CrewAI, AutoGen, Anthropic SDK, OpenAI SDK, and MCP — call instrument() once and all framework calls emit spans automatically
• Pre-built Grafana dashboard definitions for agent throughput, cost trends, error rates, and drift alerts

Quick Start

Install from PyPI:

pip install agent-observability

Verify the installation:

agent-observability version

Basic usage:

import agent_observability

# See examples/01_quickstart.py for a working example

Documentation

• Architecture
• Contributing
• Changelog
• Examples

Enterprise Upgrade

For production deployments requiring SLA-backed support and advanced integrations, contact the maintainers or see the commercial extensions documentation.

Contributing

Contributions are welcome. Please read CONTRIBUTING.md before opening a pull request.

License

Apache 2.0 — see LICENSE for full terms.

Part of AumOS — open-source agent infrastructure.