I Built 3 Checks Into My AI's Slack Workflow. I've Used One of Them Zero Times.

Short summary

The author gave an AI agent Slack posting access via MCP tools and built three verification checks into the workflow. After auditing logs, one check had never fired, and a trace the author initially described as successful actually returned zero correct results across seven search calls. The piece is a candid lesson in why 'it worked' and 'I can verify what it did' are fundamentally different claims about AI actions.

• •AI agent Slack access via MCP tools required human-verifiable checks like channel name search instead of raw ID lookup
• •Log audit revealed one of three checks never fired and a trace originally described as successful actually found nothing
• •Honest reflection on the gap between trusting AI summaries and verifying raw tool-call results

Generated with AI, which can make mistakes.