◆ Autonomous Worker
Handles address changes, stop-payments, statement copies, fee reversals and transfer setups against core-banking state. Resolves the request end-to-end and routes only the actions policy flags to an oversight agent to gate.
Memory
Working The request + the account state it touches.
Episodic Recent servicing history (repeat fee-reversal abuse detection).
Semantic Servicing policy: what's self-serve, what needs approval, fee waiver rules.
Store Core-banking state + servicing-policy retrieval
Orchestration
pipeline MCP
Harness · Managed Agents … short session per request; deterministic tool calls.
Tools
{ } Core banking servicing API API ⌕ Servicing policy (RAG) Retrieval { } Statement + document store API ⇄ Oversight-agent escalation A2A
Evals & guardrails
- Hard cap on cumulative fee reversals before forced escalation (abuse guardrail).
- Monetary-impact threshold above which an oversight agent re-derives and approves.
- Customer-confirmation step before any irreversible maintenance.
Frontier edge
- ▲Formal action-gating: every maintenance action is policy-checked and cryptographically signed before it touches the ledger, so it provably cannot exceed servicing authority and is fully replayable for audit.
- ▲Continual learning (eval-gated): online updates to its fee-reversal-abuse signals from each request, no full retrain, so repeat-abuse patterns are caught the day they emerge.
- ▲Confidential compute: runs against account state inside a secure enclave, so balances and PII are never in the clear on the servicing path.
In numbers
240,000
Servicing requests / day
88%
Self-service resolution