Harness Guardrails for Regulated Agents

Banking and healthcare cannot ship an unbounded agent. In these sectors the harness is not a convenience — it is where compliance, privacy, and control are enforced. A frontier model with a clever prompt is a liability; a governed runtime is the only thing a regulator can inspect.

The guardrails are not hoped for in the prompt. They are enforced in the runtime.

What the harness enforces in banking

Concern	Harness responsibility
Access	Least-privilege access to core banking and payment rails
Checks	KYC/AML checks; SOX and model-risk (SR 11-7) audit trails
Guardrails	Fraud actions, limits, dual-approval on money movement
Delivers	Reproducible decisions regulators can inspect

What the harness enforces in healthcare

Concern	Harness responsibility
Access	PHI minimization and HIPAA-safe data handling
Checks	EHR/FHIR access with role-based scoping and consent
Guardrails	No clinical diagnosis without human sign-off
Delivers	Full audit trail for safety and compliance review

The pattern is identical across both: scope access, gate high-risk actions on a human, and log everything reproducibly.

Anchor example — a $240 dispute and refund

A customer asks a bank agent to "dispute a $240 charge and refund me." Without a harness the agent calls a refund API directly — no identity check, no limit check, no trace. With a harness, the same request runs a governed sequence:

// 1. Identity + entitlement before anything else
verify(ctx.identity);
guard.checkScopes(ctx.identity, ["disputes:write", "payments:write"]);

// 2. Run fraud + policy tools BEFORE money moves
const fraud  = await tools.fraudCheck(dispute);
const policy = await tools.disputePolicy(dispute);   // citable, not hallucinated
if (fraud.risk > THRESHOLD || !policy.eligible) return declineWithReason(policy);

// 3. Blast-radius limit → route large refunds to a human
if (dispute.amountUsd > 100) {
  await approvals.request(ctx, "issue_refund", dispute);   // dual-approval
  return pendingHumanApproval(dispute);
}

// 4. Act, validate, and record — reproducible for audit
const out = await tools.issueRefund(dispute, ctx.identity);
audit.write({ action: "issue_refund", dispute, out, identity: ctx.identity });

Every property a regulator cares about — identity verification, policy citation, a dollar limit, dual-approval, and a reproducible trail — is a line of harness code, not a hoped-for model behavior. On error the runtime retries, validates outputs, and fails safe rather than looping or acting blind.

Where harnessed agents create value

The same guardrail pattern unlocks real work across both sectors:

Sector	Use case	What the harness guarantees
Banking	Customer service agent	Every account action permission-checked and logged
Banking	Loan & credit ops	Gathers docs, runs policy checks, escalates edge cases to a human underwriter
Banking	Compliance & AML	Triages alerts and cites evidence; final SAR call left to an analyst
Healthcare	Prior authorization	Assembles evidence against payer rules; clinician reviews before submit
Healthcare	Clinical documentation	Drafts notes scoped to PHI minimums, with provider sign-off enforced
Healthcare	Patient navigation	Handles scheduling and refills; hands off the moment clinical judgment is needed

In every case automation handles the volume and a human owns the judgment call — enforced by the harness, not left to the model's discretion.

The path to implement

1. Resolve and verify identity at the edge; attach least-privilege scopes that map to your core systems.
2. Gate every side effect behind a permission + limit check — payment rails and EHR/FHIR access are never directly callable.
3. Set blast-radius limits and route anything above them (dollar thresholds, clinical diagnosis) to human approval.
4. Run fraud, policy, and consent tools before the action, so decisions are citable rather than hallucinated.
5. Write a reproducible audit trail for every step — the evidence base for SR 11-7, SOX, HIPAA, and incident review.

GovPilot enforces access, PII handling, and approval gates as continuous agents; AI Harness provides the reliable runtime around them. Together they make an agent a regulator can inspect. The broader human-in-the-loop spine is what keeps every override defensible.

Talk to us about governed agents →