Check LangChain agent loops before they burn more tokens.

The problem

LangChain-style agents can call tools repeatedly, hit recursion limits, retry after tool errors, or keep revisiting the same structural state. Trace viewers show each call; CAUM looks at the shape across calls.

Use CAUM when you need to know whether a recurring LangChain workflow is structurally repeating, not whether the final answer is correct.

Useful event fields

You do not need prompts, source code, documents, customer messages, or private payloads. Structural fields are enough for a first check.

• tool call name
• tool result status
• phase or node label
• state_id or step fingerprint
• tokens/cost if available

Example shape

[{"event":"tool_call","tool":"retriever","phase":"plan","status":"ok","state_id":"same_query"},
 {"event":"tool_call","tool":"calculator","phase":"act","status":"error","state_id":"same_error"},
 {"event":"tool_call","tool":"retriever","phase":"plan","status":"ok","state_id":"same_query"}]

What the free check returns

• events analyzed
• exact cycle coverage
• structural pattern cycle coverage
• work-yield proxy
• burn proxy
• trace quality