SecureBank AI

Live Activity

Offline

Connect to see live events

SecureBank AI demonstrates a secure, four-agent banking assistant built with the Lauren AI framework.

Three demo users — Alice, Bob, and Charlie — each have their own account, conversation history, and real-time event stream.

Multi-agentFour specialist agents: Public CRM (pre-login), Auth CRM (logged-in), Transfer, and Disputes. The three authenticated agents form a routing triangle — each can hand off to either of the other two.
HandoffTo toolAgents hand off via a typed HandoffTo tool. The LLM receives a JSON schema with a Literal enum of valid target names — it can only route to agents explicitly wired at startup, never to an arbitrary string. All turns stream as a single SSE response; a break event marks each agent switch.
Isolated memoryEach agent has its own InMemoryConversationStore. A shared store would expose the full cross-agent history to every agent, causing them to misread prior handoff summaries as new instructions and trigger the wrong handoff.
SecureRequests are HMAC-SHA256 signed by the Next.js proxy — the browser never touches the secret key.
Real-timeA WebSocket stream pushes live tool calls, token usage, and balance changes to the browser as they happen.
Human-in-the-loopTransfers require explicit user approval before executing. ApprovalTool blocks the agentic loop with an asyncio.Future, sends a transfer_approval_request event over WebSocket, and waits (up to 30 s) for the user to click Confirm or Cancel in the approval dialog. The Transfer Agent only proceeds if a signed one-shot approval token is present in AgentContext.metadata; the token is consumed on first use, preventing replay.
Knowledge BaseThe Public CRM Agent answers product / rate / fee / branch-hour questions via RAG over a curated knowledge base, not training-data confabulation. Wired through AgentModule.for_root(knowledge=…) — the framework auto-attaches a search_public_info tool to the agent's schema. Every retrieval shows up in the Live Activity panel and as a persistent “📚 Looked up bank info” bubble in the chat thread.

SecureBank AIMulti-agent banking demo

HMAC-Signed PayloadsIdentity GuardsCRM + Transfer AgentsLive WebSocket Events

Live Activity

Offline

Connect to see live events

SecureBank AI demonstrates a secure, four-agent banking assistant built with the Lauren AI framework.

Three demo users — Alice, Bob, and Charlie — each have their own account, conversation history, and real-time event stream.

Multi-agentFour specialist agents: Public CRM (pre-login), Auth CRM (logged-in), Transfer, and Disputes. The three authenticated agents form a routing triangle — each can hand off to either of the other two.
HandoffTo toolAgents hand off via a typed HandoffTo tool. The LLM receives a JSON schema with a Literal enum of valid target names — it can only route to agents explicitly wired at startup, never to an arbitrary string. All turns stream as a single SSE response; a break event marks each agent switch.
Isolated memoryEach agent has its own InMemoryConversationStore. A shared store would expose the full cross-agent history to every agent, causing them to misread prior handoff summaries as new instructions and trigger the wrong handoff.
SecureRequests are HMAC-SHA256 signed by the Next.js proxy — the browser never touches the secret key.
Real-timeA WebSocket stream pushes live tool calls, token usage, and balance changes to the browser as they happen.
Human-in-the-loopTransfers require explicit user approval before executing. ApprovalTool blocks the agentic loop with an asyncio.Future, sends a transfer_approval_request event over WebSocket, and waits (up to 30 s) for the user to click Confirm or Cancel in the approval dialog. The Transfer Agent only proceeds if a signed one-shot approval token is present in AgentContext.metadata; the token is consumed on first use, preventing replay.
Knowledge BaseThe Public CRM Agent answers product / rate / fee / branch-hour questions via RAG over a curated knowledge base, not training-data confabulation. Wired through AgentModule.for_root(knowledge=…) — the framework auto-attaches a search_public_info tool to the agent's schema. Every retrieval shows up in the Live Activity panel and as a persistent “📚 Looked up bank info” bubble in the chat thread.

Banking Assistant

Public (Guest) · not logged in

Talking to: Banking CRM Agent (Public)

Hi! How can I help you today?

Payloads signed with HMAC-SHA256 · Log in for account access