agentSonar

agentSonar is a monitoring and observability tool designed specifically for AI agent systems, with a focus on detecting silent failures and inefficiencies that traditional tracing tools often miss. It addresses issues such as undetected loops, runaway costs, and coordination failures within agentic workflows by modeling the interactions between agents rather than just individual execution spans. The platform is intended for developers and teams working with complex AI agent orchestrations, particularly those using frameworks like LangGraph, CrewAI, and Claude Code.

A core feature of agentSonar is its ability to catch eight classes of silent failures that do not throw exceptions or generate error logs. These include silent loops where agents pass work in circles, repeated agent calls with no progress, unexpected traffic spikes, redundant tool calls, stuck tool calls, subagent explosions, failed-tool retry storms, and context-window cliffs where session context nears its limit. Upcoming features are planned for detecting deadlocks, projecting cost runaways during execution, and identifying ungrounded responses.

agentSonar offers two operational modes: Detect Mode, which streams real-time alerts to stderr as issues are detected, and Prevent Mode, which raises a typed PreventError before the next language model call to halt problematic runs and save costs. After each run, the tool generates a standalone HTML report, along with an alerts log, a JSON report, and a JSONL timeline. These reports require no external dependencies and can be shared or archived as needed. The tool is available as a Python package (installable via pip) and a Node/TypeScript package (installable via npm), with quick integration into supported frameworks through simple configuration steps. A quick demo is provided for immediate testing without API keys or remote services.

0 open-source license and is currently in a closed beta phase. It is positioned within the class of AI agent monitoring and observability solutions, offering advanced detection and prevention capabilities for agentic workflow failures.