Introduction

BankGPT is the backend integration layer that gives an AI agent hands inside applications with no API.

An LLM discovers a back-office UI flow once — driving a real browser, observing the accessibility tree, one structured model call per step — and the successful run is distilled into a typed, versioned capability artifact. From then on the flow replays deterministically with zero model calls: a calling AI agent invokes the capability with typed inputs and gets back typed outputs, a known business outcome, or a debuggable failure. Human operators approve risky actions (with maker ≠ checker segregation) and can take over the same live session when automation gets stuck, then hand control back.

The through-line:

goal → discovery (model in the loop) → artifact (typed contract)
     → replay (no model) → structured result

with policy enforced and evidence captured at every step.

The system at a glance

Why it works this way

• Discovery once, replay forever. Model calls are expensive, nondeterministic, and unauditable. They happen exactly once per flow; the recorded artifact is what runs in production.
• The artifact is the contract. Human reviewers read it, the replay executor interprets it, and the calling agent treats it as a typed function signature. One zod schema is the single source of truth.
• Humans stay in the loop. Risky capabilities refuse to run without a segregated operator approval, and a stuck run hands the same live browser session to a human and back — no restart, no hidden state.
• Every run leaves evidence. Step logs, model transcripts, failure screenshots, and the control log are stored per run and served back over the engine API.

Where to next