Guide

Regulated Agent Harness Architecture

A practical architecture for agent authority, tool and data scope, pre-action policy, validation, human escalation, monitoring and evidence.

For financial-services platform architects, AI risk leads and security teams.

Last updated: Sep 7, 2026 · Version v1.0 · Not legal advice.

Download the one-page architecture (PDF, English) Discuss your agent control architecture

Short answer

A regulated-agent harness connects an agent’s proposed action to explicit authority, bounded access, policy evaluation, validation and human escalation, then records the execution outcome. KLA’s reference architecture organizes that work into seven controls.

Architecture

The action path

Business owner grants authority → agent proposes action → identity and scope checks → pre-action policy → validation or human decision where required → controlled execution → outcome and evidence.

Apply the sequence to each consequential action. Establish network and credential boundaries around it, and send operational failures to the monitoring and remediation process.

Control map

Seven controls and their owners

KLA’s product mapping below describes relevant surfaces. The acceptance column is a deployment test to perform; it does not assert estate-wide enforcement or completed customer verification.

Control	KLA mapping	Accountable owner	Acceptance test
Authority	Agent Registry; governed request identity	Platform and business owner	Unrecognized or out-of-scope agent cannot perform the action
Tool and data scope	Tool Catalog; Data Boundaries	Platform and data owner	Restricted tool, record and tenant requests are rejected
Pre-action policy	Policy Builder; KLA Policy Engine	Policy owner	Exercise allow, warn, require_approval and block on the integrated path
Validation	Simulation; workflow-specific checks	Independent reviewer	Invalid output and unsafe state transitions fail before release
Human escalation	Decision Desk	Authorized decision owner	Rejection and expiry prevent the approved-action path from executing
Monitoring	Assurance Center; Lineage Explorer	Operations and security	A control failure becomes an owned finding with a response
Evidence	Audit Trail; Evidence Room	Evidence and records owner	Reconstruct request, decision, human review and actual execution outcome

Integration

Define the enforcement boundary

List every tool endpoint, credential, data source and delegated agent that can cause the selected action. Route the governed action through the checkpoint and test alternate paths. Record any path that remains outside enforcement.

The host environment owns sandboxing, network isolation and production access. An approval must apply to the concrete action and remain valid at execution; specify behavior when parameters, policy or authority change.

Acceptance

Review one case end to end

Use a synthetic case with a permitted read, an action requiring approval and a blocked operation. Check downstream state after approval, rejection, expiry and retry. Record identifiers, policy version, reviewer identity and result.

For sealed exports, also run the independent verifier and retain its result. A visible record and a successfully verified bundle provide different evidence; label each accurately.

Context

Source and interpretation

This is KLA’s architecture proposal, informed by the Bank of England’s harness note. Its seven controls are KLA’s grouping. The accompanying article explains the Bank’s six themes and the note’s scope.

Map the controls to EU AI Act standards topics Download the one-page architecture (PDF, English)

FAQ

Questions buyers should resolve

Does a harness replace model evaluation?

Model evaluation remains part of system validation. A harness also needs tests of authority, access, action execution, human decisions and operating failures.

What does KLA control in this architecture?

KLA contributes policy, human decisions and evidence for configured governed paths. Verify integration coverage and separately assign hosting, network, credential and system-level assessment responsibilities.

Links

Sources

Bank of England: Frontier AI harness engineering