AgentOps is an observability and developer-tools platform for AI agents. Its Python SDK instruments your agents in a couple of lines of code and records what they do — sessions, LLM calls, costs, tool use and failures — into a dashboard where sessions can be replayed and debugged. The project has about 5,900 stars on GitHub, is MIT licensed, and was actively developed in 2026.
Agents fail in ways that are hard to see: a loop here, a hallucinated tool call there, a quietly ballooning token bill. AgentOps gives teams the missing visibility layer. The SDK integrates with the major agent frameworks — CrewAI, Agno, the OpenAI Agents SDK, LangChain, AutoGen, AG2, CAMEL-AI and others — and with most LLM providers, tracking cost per session and per call, benchmarking runs, and surfacing errors with full event traces. Data lands in the hosted dashboard at app.agentops.ai, where sessions can be inspected and replayed step by step.
AgentOps is frequently cited as the observability layer in agent tutorials and framework documentation, and its two-lines-of-code pitch is close to literal: initialize the SDK with an API key, instrument your agent, and calls start appearing in the dashboard. The session replay is the feature teams usually point to — stepping through exactly which prompt, tool call and response led to a bad outcome — since agent bugs are otherwise nearly impossible to reproduce from plain logs alone.
Freemium. The AgentOps SDK is free and open source under the MIT license; the hosted platform offers a free tier, with paid plans for teams that need more volume and collaboration features.
Teams building and shipping AI agents who need to see what their agents actually did, what it cost, and why a run failed.