The architecture behind a trustworthy finance agent

People assume the model is the product. In a finance agent, the model is the smallest layer in a stack that exists almost entirely to make that model safe, accurate, and accountable. Here is the architecture, roughly from the outside in, and why each layer earns its place.

The control layer

Wrapping everything is the control layer: the guardrails that decide what runs autonomously, the approval flows that gate every financial action, and the role-based access that enforces least privilege. This is the layer finance cares about most, and it sits outside the model precisely so that no model behavior can bypass it.

The evidence layer

Underneath every action is the audit trail: an append-only record of who did what, when, and on what basis, written at the moment of action. It is not a log you can disable; it is structural, so the system cannot take an action without recording it. Evidence is a property of the architecture, not a feature you remember to turn on.

The orchestration and tool layers

• Orchestration breaks a goal into steps, sequences them, and checks results.
• Tools give the agent governed access to your systems under scoped permissions.
• Verification re-checks high-stakes outputs before they are presented for approval.

The model, finally

At the center sits the model — or models, since different tasks call for different ones. It does the reading, the extraction, the drafting. But it never acts directly on your ledger; it proposes, and the layers around it verify, gate, approve, and record. That separation is the whole point. The intelligence is useful; the structure around it is what makes the intelligence trustworthy.

When you evaluate a finance agent, you are really evaluating these outer layers. The model is increasingly a commodity. The trustworthy system around it is not.