The Bank of England’s 2 September 2026 technical note on harness engineering puts the surrounding engineering of frontier AI at the centre of practical deployment. Its focus is cyber defence. It reports forum discussions with systemically important UK financial institutions, the FCA, HM Treasury and NCSC, and creates no new supervisory expectations.
For teams building financial-services agents, our architectural reading is straightforward: review the complete path from delegated authority to an external action. That is the point at which model capability becomes an operational responsibility.
What is an AI harness?
An AI harness is the software and operating environment that lets a model do work: context, tools, execution flow and controls. For a regulated agent, KLA’s design adds an explicit record of who may authorize an action, which boundary enforces that authority, and what evidence survives the run.
Six themes, translated into architecture decisions
The following short labels summarize the Bank’s six themes. The design questions in the second column are KLA’s interpretation for regulated-agent architecture.
| Bank theme, summarized | KLA architecture question | KLA mapping |
|---|---|---|
| Harness design | Where is each consequential action checked? | KLA Policy Engine; Decision Desk |
| Component choices | Who owns each control across internal and supplier components? | Tool Catalog; Provider Hub |
| Orchestration | Does authority remain bounded when work moves between agents? | Processes; Agent Registry |
| Sensitive context | Which tools, records and environments may each agent access? | Data Boundaries; Tool Catalog |
| Embedded controls | Which decisions block execution or require a human? | Policy Builder; Decision Desk |
| Validation and scale | Can reviewers validate findings and close remediation at the expected volume? | Simulation; Assurance Center; Evidence Room |
Start with the action that changes something
Consider an illustrative security agent that finds a dependency vulnerability in a bank’s application repository. Reading an approved source snapshot, proposing a patch and deploying the patch carry different authority. Define separate permissions for each action and assign the production change to the bank’s change owner.
A useful acceptance demonstration has the agent propose a valid patch, request deployment outside its mandate, and encounter a rejected approval. Inspect the destination system after each attempt. A policy decision record alone cannot establish that a downstream write was prevented.
Where KLA fits
KLA Control Plane provides policy and decision records around configured governed action paths. Policy Builder, the KLA Policy Engine and Decision Desk connect policy authoring, action evaluation and human review. Tool Catalog and Data Boundaries describe governed access; Lineage Explorer and Evidence Room support investigation and evidence review.
Coverage depends on the integration. Inventory direct SDK calls, secondary credentials, delegated agents and network routes. An external agent that retains an unrestricted alternate execution path can bypass a checkpoint. Network isolation and credential custody also need controls in the host environment. KLA’s presence in a diagram does not establish those boundaries.
For a concrete mapping, use the Regulated Agent Harness Architecture. It assigns each control an owner and an acceptance test.
A model change should trigger a harness review
When a team replaces a model, review its available tools, retry behavior, input handling, delegation and failure recovery. Re-run the blocked-action and approval tests with the new setup. Record model, tool and policy versions so a later investigation can identify the configuration that produced the action.
Track review workload as well as model quality: unresolved findings, age of pending approvals, repeated requests and remediation completion. A system that generates more findings than the team can resolve needs a change to scope, prioritization or operating capacity.
Connect the architecture to standards preparation
Our EU AI Act control mapping connects the same design work to risk management, cybersecurity and logging topics. It is an engineering preparation aid. Applicability and conformity require assessment of the particular AI system, its legal role, the relevant requirements and the standards actually applied.
The Bank note offers a useful reference for architecture discussions. It does not endorse KLA or establish a requirement to purchase a control plane.
Frequently Asked Questions
Is the Bank of England harness note new regulation?
The note reports technical discussions and creates no new supervisory expectations. Institutions still need to assess their existing obligations and their own deployment.
How should a bank evaluate an agent harness?
Choose a consequential action, define its authority and access boundaries, test rejection and human escalation, inspect the downstream effect, and retain the decision and execution evidence.