Your LangGraph agent ran 50 iterations on what should have been 3.

Recursion edge cases, tool-call retries, and self-correcting loops are the most expensive way to fail an LLM workflow. Calcis caps the loop before the bill catches up: a per-route iteration ceiling, a per-PR predicted-spend gate, and an inline VS Code warning when an agent's projected cost crosses the line you set. Longer write-up plus a CI recipe landing here next.