Banking

What the bank is asking for

A quarterly review names four hundred corporate accounts. For each one, someone opens the customer file, the transaction history and the current regulatory requirement, decides whether the review is complete, and writes down why. The list is right; the reasons take three weeks.

• List the accounts whose review is outstanding, with the reason for each and the record that proves it.
• Follow an unusual pattern of transactions through the systems that hold it, and write what was found.
• Compare a new product against the disclosure requirement that applies to it, clause by clause.
• Answer a branch question from the bank’s own procedures rather than from memory.

How the work is done

An agent is a model, a set of approved skills, tools and a policy, applied to a scope.

Agents answer from your own regulations, documents and records, with the source attached.

Agents reach your systems through approved connectors — APIs, databases, applications — each with its own permissions.

Changes wait for human approval before they run.

Customer-controlled inference and execution boundaries.

Before any of that touches customer data, the bank’s own infrastructure is the safer first workload: the servers behind the channels, the batch jobs and the branch estate. Those questions are how a bank tests whether the controls hold.

Read-only execution policy is the default on every enrolled host; the full policy is an explicit, per-host choice.

Every answer shows the exact commands that ran, their exit codes and output, per host.

A host joins the fleet with one install command and an enrollment token; revoking the host unpairs it.

Sign in with your own identity provider (OIDC with PKCE); no shipped default administrator password.

Bring your own model — OpenAI-compatible endpoints, Ollama, Anthropic, or Cloudflare AI Gateway / Workers AI — with keys envelope-encrypted at rest.