Assisted service and agents

Nucleus’s agentic banking target includes institution, customer, business, partner and platform agents. Each acts through a current principal grant, with ordinary domain commands, permissions, budgets and approval controls.

Assisted human service

Staff first verify the caller or representative, use the least customer information needed and act only within the assigned service scope. Retain the case, verified instruction, decisions, financial record and communication status. A disconnected call leaves owned work that another authorised staff member can recover.

Agency banking has a different actor model involving institution agents, devices, float and physical custody. The core /agents resource describes those banking-network actors.

Software agent execution

The existing AI gateway exposes catalogue, ask, query and audit interfaces with source-specific limitations. Its sample/query capability grants no financial execution rights.

The proposed agent contracts describe registered agents, scoped tools, versioned mandates, exact prepared actions, independent decisions, atomic budgets and fenced execution. Qualify each actual command and institution configuration before enabling financial execution.

Prepare the workflow

The accountable principal and agent owner select a local or sandbox environment, allowed accounts and beneficiaries, tool/command versions, purpose, time bounds, budgets and recovery owner. Required human approval binds the exact action version and hash. Revocation prevents new execution while already dispatched obligations continue through recovery.

The interface shows what is proposed, the principal and grant, the required decision, execution state and how an authorised person can stop new work or take over recovery. A natural-language response cannot establish permission or settlement.

Record acceptance

Exercise changed payloads after approval, concurrent budget use, revocation before and after dispatch, expired grants, exhausted runtime budget and interruption. Retain authority decisions, operation and posting identities, external evidence and the recovery handoff. Agent-specific compute limits must leave essential reconciliation and recovery services able to operate.