agentsop-llm-tool-idempotency

To prevent duplicate side effects from LLM agent tool retries, implement at-most-once execution using a caller-supplied idempotency key and deduplication at the execution boundary. This ensures retried calls become safe no-ops instead of triggering duplicate external actions.

Duplicate tool calls in LangGraph happen due to framework resume, transport timeouts, or model duplicate emissions. Make them idempotent by implementing execution-boundary dedup logic with stable idempotency keys, ensuring the second call is blocked as a no-op with observability.

The best patterns for idempotency in distributed LLM systems include native idempotency APIs with headers, internal dedup tables, content-addressed writes, and database upsert patterns. These approaches block duplicates at the execution boundary rather than silently re-triggering effects.

Yes, you can use idempotency keys to handle MCP transport timeouts safely. By requiring caller-supplied stable keys and applying execution-boundary dedup logic, retried MCP tool calls become no-ops, preventing double charges or duplicate records when timeouts occur.

When an LLM agent sends the same idempotency key with different arguments, conflict handling rules apply. The protocol includes production guardrails and anti-patterns to ensure duplicates are blocked, preventing inconsistent state from mismatched retry arguments.

You do not need native API support to deduplicate tool calls in LLM agents. The protocol covers framework-agnostic operating models including internal dedup tables, saga compensation, and orchestrator-specific guidance for LangGraph interrupts and MCP tool schemas.