The original title is "Building an AI Agent That Knows When Not to Guess (Qwen + MCP)"

Short summary

The author built Recona, an AI agent that reconciles Paystack payments against open invoices using Qwen and MCP, exposing matching and reminder-drafting as MCP tools over streamable HTTP. During real testing, Qwen correctly identified the right invoice in its reasoning but returned only 30% confidence and refused to commit—demonstrating honest self-uncertainty. The key design lesson: build guardrails around the model's own confidence output rather than trying to outsmart the model at judging its own answers.

• •Recona uses Qwen + MCP to auto-reconcile Paystack payments and chase overdue invoices with deterministic guardrails
• •Qwen correctly reasoned about the right invoice but returned 30% confidence and declined to commit—honest uncertainty as a usable signal
• •Design principle: treat model confidence as a first-class output and gate auto-actions on it, rather than trying to judge the model's answers yourself

Generated with AI, which can make mistakes.