Structured Data Extraction
Your structured data extraction pipeline uses an MCP tool named mcp__crm__write_record to persist extracted contract fields back to the CRM. After deploying, you observe that roughly 12% of write operations are immediately retried by the agent — logs show the agent receiving a deny response but then re-attempting the same write within the same session. The PreToolUse hook correctly identifies and blocks writes containing invalid contract_id formats, but the agent keeps retrying. What is the most effective fix?