AgentTelemetry is an open-source observability framework for autonomous AI agent systems built on large language models. It introduces agent-specific semantic conventions for distributed tracing — including span kinds for reasoning, LLM calls, tool invocations, planning, retrieval, and inter-agent communication — that extend OpenTelemetry for agent workloads. The framework provides auto-instrumentation for LangChain, CrewAI, AutoGen, and DSPy, W3C-compatible context propagation for multi-agent tracing, privacy-first design with opt-in content capture, built-in cost tracking, and export to any OTLP-compatible backend. Code and tests are available at https://github.com/Krishnachaitanyakc/AgentTelemetry.