LLM Agent Crash Recovery and Retry Deduplication Tested Across 3 Frameworks in 34 Runs

A developer ran 34 controlled experiments across LangGraph, Strands, and CrewAI to measure how AI agents handle process crashes and duplicate side effects during LLM retries. LangGraph with a durable checkpointer recovered from a SIGKILL in under 0.02 seconds with zero additional LLM calls, while Strands and CrewAI performed full re-runs averaging over 5 and 2 LLM calls respectively. The study also tested three idempotency key strategies to prevent duplicate tool executions on retry, finding that content-hash keys silently fail when the LLM rewords its arguments during a retry, producing a different hash and bypassing deduplication. Position-based keys, which identify a tool call by its place in the workflow rather than its content, proved more reliable at catching duplicates across retry scenarios. The findings highlight that true crash recovery requires state stored outside the process, and that LLM retries cannot be treated like standard HTTP replays due to non-deterministic argument generation.