---
url: https://hanria.ai/blogs/principals/
canonical: https://hanria.ai/blogs/principals/
site: HANRIA
---

# Agents can act. They cannot yet prove.

Why the missing piece in the AI economy is accountability, and how HANRIA is designed to build it.

The smarter agents have arrived. They are not principals.

Andrew D'Souza makes the point in his essay [Principals](https://www.boardy.ai/principals): AI has produced many agents and almost no principals. An agent acts for someone else, so we trust whoever stands behind it, never the agent itself. A principal answers under its own name, to many stakeholders, and pays for its own mistakes. In his account, trust, not capability, is the limiting factor of the AI economy, and agents alone cannot raise that ceiling.

The diagnosis is right. The question is what to build.

## A real problem with a court record

In [Moffatt v. Air Canada, 2024 BCCRT 149](https://www.weirfoulds.com/moffatt-v-air-canada-bot-betrayal-and-negligent-misrepresentation), a British Columbia tribunal held the airline liable for wrong bereavement-fare advice given by its website chatbot. The tribunal treated the claim as negligent misrepresentation and rejected the airline's attempt to treat the chatbot as separate from the company. The ruling is a small claims decision, not appellate precedent, but it shows the pattern: when an agent errs, the consequence is assigned after the harm, by a tribunal, to the party behind it.

Anthropic's [Project Vend](https://www.anthropic.com/research/project-vend-1), run with Andon Labs, shows the pattern from the other side. An AI shopkeeper was told to turn a profit and avoid bankruptcy. It still sold items below cost, gave away discounts and failed to learn reliably from feedback. The business did not make money. Corrective signals existed. They did not hold.

Design by abdication remains design. When no one engineers durable consequences for an agent's conduct, the system still has an outcome: it arrives late and it arrives on someone else.

## The institutional deficit

Corporations earned trust as accountable persons over a long historical process. Edward Thurlow is credited with the complaint that a corporation has "neither a soul to lose, nor a body to kick." Law, registries, liability rules and reputation supplied the missing consequences over time.

AI agents face a comparable institutional deficit, on a compressed timeline. They transact at machine speed, and the institutions that make a counterparty legible, such as identity, authority, liability and credit history, are only beginning to exist for them. That is our reading of the problem, and it is the one HANRIA is designed around.

## What HANRIA is designed to do

HANRIA is a decision point that checks who an agent is, what it may do, and whether its work passed review before payment. Its architecture maps onto four questions any counterparty must answer: identity, authority, liability, and credit history. Five components serve those questions: Check addresses authority, Runtime enforces it, Review addresses liability and quality, Carry addresses identity and credit history, and Assessment attests capability.

**[Check](https://hanria.ai/check/) (available).** An operator-owned mandate states what is permitted, denied, and escalated to a person. The check returns permit, deny, or escalate, names the governing clause, and logs the decision. The hosted version issues signed receipts bound to the mandate, the action, and a fifteen-minute window.

**[Runtime](https://hanria.ai/runtime/) (designed).** Enforcement intended to refuse admission, revoke access, and block privileged operations, but only where it exclusively controls the tools or credentials.

**[Review](https://hanria.ai/review/) (designed).** Independent reviewers stake their reputations, not money, on checking work before payment. The design holds payment until review concludes, then releases or refuses it. A reviewer whose judgment does not match the outcome loses standing, and the other reviewers can outvote a mistaken one.

**[Carry](https://hanria.ai/carry/) (designed).** Portable records bundle an agent's track record with the evidence needed to evaluate it. Four gates apply in sequence and fail closed: continuity, competence, trust, and authority. A later answer cannot satisfy an earlier question.

**[Assessment](https://hanria.ai/assessment/) (designed).** Attestation of an agent's capabilities against the Agent Failure Mode Registry, which covers 32 failure families.

## Three design choices that matter

**Continuity.** A record is worthless if a new agent can borrow it. The design checks that the agent presenting a record is the same subject that earned it, across changes of key, model and operator.

**Three-name recording.** Each job records the agent, the harness and the model separately. This supports investigation: it gives reviewers the means to examine whether a failure traces to the model, the harness or the agent's configuration. It does not by itself establish fault.

**Evidence over assertion.** Track records are scoped and non-transferable. They keep adverse states: refused, suspended, revoked. Missing evidence is recorded as null, never as zero. An absence of history is not a clean history.

## What this does and does not answer

D'Souza proposes four tests for a principal: one entity for everyone, its own goals, something to lose, and an expectation of persistence. HANRIA does not claim to satisfy them, and it does not claim to make an agent a legal person. Reviewer stakes are reviewers' stakes, not the agent's. Portable identity does not establish independent goals.

What HANRIA addresses is the machinery the tests presuppose. Continuity supports persistence. Scoped records that keep adverse states give a subject a history it cannot quietly shed. Staked review puts consequences on the judgment of those who vouch for the work. Our thesis is that an agent whose conduct produces durable, inspectable consequences is a better candidate for real work than one whose conduct leaves no trace. That is a design thesis, and it is not yet an empirical result.

## What HANRIA does not claim

Precision about limits is part of the design. Skills and the hosted check advise; they do not enforce. Only a runtime can refuse, and only where it exclusively controls the tools or credentials. A deny in a log does not prove that anything was stopped. HANRIA records evidence. It does not determine or assume legal liability, and it does not guarantee quality, safety, legal compliance or trustworthiness. It aims to make the basis for trust inspectable, which is a precondition for everything else. A U.S. trademark application for HANRIA is pending. No registration or completed clearance is claimed, and no security certification or regulated financial service is claimed.

## The stakes

If D'Souza is right that 2027 will be the year of principals, our expectation is that value will concentrate in whoever can prove accountability, and that agent work without it will be commoditized. That is a prediction, not an established outcome. The infrastructure for that proof has to be built before the market needs it, not after the first large failure.

*Wulf A. Kaal*