A bank can't let a single AI approve loans on its own. It needs a committee with separated roles and a clear record of who decided what. We built exactly that: an autonomous credit committee where six AI agents collaborate through Band to originate, approve, and disburse real USDC loans on-chain. Band is the coordination layer. The six agents , a Borrower, a Gatekeeper (identity/KYA), an Analyst (credit risk), a CFO (treasury), a Settler (on-chain disbursement), and an Auditor , live in a shared Band chat room and hand off work to each other by @mentioning. Turn Band off and the flow stops: the disbursement happens because the CFO approved it in the room, not as an afterthought. Under the room sits a two-layer design. Each agent is a LangGraph ReAct agent (GPT-4o + tools) connected to Band through the band-sdk LangGraph adapter. The agents call the RSoft Agentic Bank backend (FastAPI + LangGraph + Supabase) over REST, which runs the real lending logic and executes an ERC-20 USDC transfer on Base Sepolia. Band moves the messages between agents; the bank API does the financial work; the agent is the bridge. This isn't simulated text. A full committee run disburses real USDC, verifiable on BaseScan, and the Auditor closes each case with an explainable verdict , the credit score, the approval, the settlement tx hash, the block number , exactly the audit trail that regulated finance demands. This is what becomes possible when agents from different layers can discover each other, coordinate, and execute real work together: not one AI doing everything, but a team of specialized agents running a bank.
Category tags: