Agentic ReAct loop

1. 1Read intent and build a bounded plan with an explicit budget.
2. 2Retrieve context, then call one schema-validated tool at a time.
3. 3Feed the observation back and validate it before the next step.
4. 4Check the policy gate before any write or high-impact action.
5. 5Stop on goal, block or escalation — and emit the full trace.

1 get_customer(CUST-4412)      obs: risk profile MEDIUM, KYC verified 2024-11
2 retrieve_transactions(90d)   obs: 412 rows, 3 above threshold
3 graph_traverse(beneficiary)  obs: 2-hop shell company link
4 sanctions_screen()           obs: no direct hit, 1 fuzzy 0.81
5 policy.evaluate()            DENY -> human escalation, 5 audit events written