Agent insurance needs an underwriting bill of materials

Harbor came out of stealth on 13 October 2026 with $10 million to measure, mitigate and insure AI-agent risk. The launch is interesting less because insurance has arrived than because it makes the machinery behind the policy visible.

My claim is that agent insurance needs an underwriting bill of materials. Treat the policy as a versioned build contract that names the agent, controls, evidence, broker, carrier, covered perils and exclusions. An insurance logo without that record is a safety badge attached to a moving system.

Harbor turns agent insurance into a control pipeline

Harbor's launch announcement describes a sequence: find every agent, measure operational, security and regulatory risk, block decisions that violate company rules, then insure what remains. Its product page adds monitoring, runtime guardrails, static analysis and adversarial testing. It says detections are attributable to a session and agent version, with analysis re-run when the system changes.

That last detail matters. An agent is not one durable object. Change the model, tool list, permission scope, data connection or approval threshold and you have changed the risk being presented to the insurer. A policy bought against yesterday's configuration cannot become today's assurance by inertia.

This extends the control boundary I argued for in agent approvals need independent policy checks. The policy check decides whether an action may run. The underwriting record decides which version, evidence and remaining failure modes the insurance decision actually describes.

The underwriting bill of materials records the change

This is labelled editorial judgement, not a Harbor feature I tested. For an enterprise architecture review, I would require one record containing the deployed model and prompt revision, enabled tools, permission and data scopes, enforced controls, evaluation evidence, named perils, exclusions, broker and risk-bearing carrier. Every material change should produce a diff and a decision to accept, mitigate or re-underwrite the changed risk.

Harbor's risk framework supports the shape of that record. It scopes risk by agent architecture, industry and jurisdiction, then separates technical risks from governance risks. It also names concrete controls: schema-check tool calls, isolate untrusted input, use scoped short-lived credentials, register agents before deployment and log decisions with their inputs.

The original contribution is the join between those technical fields and the policy. A security review can say a guardrail exists. An underwriting bill of materials must also say which version was assessed, which failure it addresses, who accepted the remainder and whether the issued policy covers that remainder.

Harbor's broker and carrier are different parties

The public launch says Harbor puts its own money behind the answer. The legal page is more precise. Harbor's 13 October licensing disclosure says Miralta Q, Inc., doing business as Harbor, acts as the distribution and marketing agent for insurance carriers. It holds producer and surplus-lines licences across much of the United States, while the products are issued by surplus-lines carriers.

That is not a contradiction I can resolve from the website, and it should not be papered over. Risk assessor, broker and carrier are different roles. The same disclosure says coverage depends on the actual policy's terms, conditions, exclusions and limitations, and may not be available in every jurisdiction.

So the buyer's question is not merely, “Is this agent insured?” Ask who measured the risk, who enforced the controls, who placed the policy, who bears the insured loss and which configuration the evidence covered.

Where this holds and where it breaks

I have not tested Harbor, reviewed a specimen policy, verified its customer deployments or observed a claim. The company pages establish the product design and regulatory roles; they do not establish loss reduction, coverage for a particular incident or the reliability of the platform in production.

Insurance also does not replace engineering. A policy can transfer a defined financial remainder. It cannot restore leaked data, reverse every harmful decision or make an unbounded agent acceptable. The useful move is to let underwriting force precision into the build record, then keep the control plane independent of the agent's own explanation.

What's in it for you

• Attach the policy to an exact agent version, tool set, permission scope and evidence pack.
• Record assessor, broker and carrier separately; ask for named perils and exclusions from the issued policy.
• Trigger a new risk decision whenever a model, tool, data source or control boundary changes.

Insure the residual risk, but version the system and evidence that defined the residual first.

Sources

• Harbor — AI needs clear liability, 13 October 2026.
• Harbor — Product, accessed 14 October 2026.
• Harbor — AI Agent Risks in Enterprise, accessed 14 October 2026.
• Harbor — Licenses and regulatory disclosures, 13 October 2026.