Build a Break Triage Agent & Flow

🎯 What You'll Build

A Break Triage Agent that classifies reconciliation exceptions by type and severity and drafts a disposition note for a human preparer — never clearing a break on its own — then wrap it in a Flow and connect a real, public, no-auth MCP server for a live data lookup.

🪜 Part 1 — Build the Agent

1. Create the Break Triage Agent with a natural-language brief: classify each break by type (timing, missing entry, duplicate, FX mismatch) and severity, draft a disposition note citing the source line, and never auto-clear a break.
2. Test it against a sample reconciliation export with at least 3 different break types.

🔁 Part 2 — Iterate Twice

1. Add a capability: have it also flag any break open for more than 2 consecutive cycles as requiring escalation.
2. Add a guardrail: explicitly instruct it to never estimate a severity threshold not defined in the reconciliation procedure — it must say "not defined in the procedure" instead.

📄 Part 3 — Generate a Disposition Memo

Ask the agent to produce a formatted disposition memo for one flagged break, with named sections (Break Description, Classification, Source Citation, Recommended Next Step) — confirm it never recommends "clear" as the next step.

🔁 Part 4 — Wrap in a Flow

1. Create a Flow: Agent Step (run Break Triage) → Decision Branch (escalate if severity = high) → Document Generation (disposition memo) → Human Approval Gate (preparer sign-off).
2. Schedule the Flow to run each business morning.

🔌 Part 5 — Connect a Real Public MCP

Connect a real, publicly available, no-auth MCP server (e.g., a public reference-data or documentation MCP) to give the agent a live lookup capability, and confirm the agent cites the live tool call in its output rather than a stale document value.

✅ Success Criteria

• The agent never auto-clears a break in any test
• Both iterations (capability + guardrail) are demonstrably reflected in a test run
• The disposition memo has named sections and a clear, non-approving "next step"
• The Flow includes a working Human Approval Gate
• The MCP-sourced value in agent output is distinguishable from a document-sourced value