AgentSonar
A monitoring tool that watches the traffic between AI agents and tool calls to catch coordination failures a normal tracing tool misses, such as silent loops, repeated work, runaway subagent fan-out and hung or retrying tool calls.
AgentSonar analyzes the shape of traffic between agents and tools rather than individual calls, because each call in a failing coordination pattern can look fine on its own. It detects silent loops such as agents handing work back and forth forever, agents repeating the same request many times, runaway subagent fan-out, hung tool calls, and failing-tool retry storms, and it can alert when a session is sliding toward the context-window limit.
It works with a single Claude Code session and its subagents, and with multi-agent orchestrators including CrewAI, LangGraph, and custom Python, Node or Electron orchestrators; unsupported frameworks can use a custom integration path. An HTML report draws a detected loop as a graph with its rotation count and the moment it tripped, and an optional prevent mode halts a run automatically before the next LLM call once a threshold is crossed.
It installs locally with no account and no remote dashboard, and is distributed through PyPI and npm.