Review fraud cases with multiple MongoDB Atlas signals

Goal

Ground a managed agent’s case summary in traceable Atlas signals and leave the decision to a human.

Steps

1. Use $vectorSearch for similar decided cases, $search for exact entities/IDs, and $rankFusion to combine rankings. [mongo]
2. Run $graphLookup over linked accounts; keep retrieval tools read-only and return transaction IDs, signal types, and evidence. [mongo]
3. Have the agent propose approve/reject; pause risky cases at requires_action, then let the application write the decision and append-only audit record after review. [mongo]

How to verify

Vector, lexical, and relationship evidence link back to records; risky decisions pause for review; decisions and audit entries can be read back for this run.

Troubleshooting

If the Atlas version lacks `$rankFusion`, fuse rankings in the application; keep MongoDB credentials in the host handler, outside agent context and sandbox.