The Missing Strand
When entities are operated by autonomous agents, limited liability depends on a layer that has not been built yet: evidence of what the agent actually did.
A sitting government has proposed a legal category for entities operated by autonomous AI systems — companies that run without human employees or managers, holding limited liability of their own. The framing reaches back four centuries: as the limited-liability company once unlocked modern capitalism, a new legal form is offered to unlock the agent economy. The machine and the legal entity, paired, as the double helix of a new kind of prosperity.
It is a serious idea, and the historical parallel is apt. But the parallel carries a detail its champions tend to pass over.
Limited liability did not work because of the legal form alone. It worked because the legal form arrived alongside something else: books. Ledgers, records, audited accounts — a contemporaneous trail of what the entity did, who it transacted with, and what happened when. Liability could be limited precisely because it could first be determined. The shareholder’s exposure had a ceiling because the entity’s conduct had a record. Remove the record, and “limited liability” describes nothing — there is no way to establish what one is being held liable for, or relieved of.
This is the strand the agent-economy proposals leave implicit.
When an entity is operated by a human, the record is a byproduct of how humans work: emails, approvals, signatures, minutes. When an entity is operated by an autonomous agent, that byproduct does not exist by default. The agent reasons, retrieves, calls tools, and acts — and unless something is deliberately built to observe and retain that sequence, it leaves no contemporaneous trail an outside party can later read.
The consequence is concrete. The day something goes wrong — a transaction that shouldn’t have happened, a counterparty harmed, a regulator asking — the question is not who is liable. The legal form may answer that. The question is the one underneath: what did the agent actually do? And if the answer depends on reconstructing it after the fact, from whatever happened to be retained, the liability — limited or otherwise — rests on sand.
This is the distinction worth naming. There are two layers, and they are not the same.
The legal layer defines who is responsible. It is moving quickly; governments are competing to define it.
The evidence layer defines how one proves what occurred. It is moving slowly, and in the case of autonomous agents, it largely does not yet exist.
A legal form that assigns responsibility without an evidence layer beneath it is a ceiling with no floor. It tells you the limit of liability without giving you the means to establish the fact. The two were always meant to arrive together. In the original double helix, the entity and the record were inseparable; one without the other was not prosperity, just exposure.
The agent economy will need its equivalent of the ledger — audited traceability of what the agent did, produced independently of the party operating it, durable enough to hold when scrutiny arrives. Not as a feature. As the precondition that makes the rest coherent.
The legal imagination is racing ahead. The evidentiary imagination has not yet left the station. Whoever is serious about the first will, sooner than they expect, have to be serious about the second.
Receiver-side behavioral observation. Evidence, not enforcement. References editorial and legislative developments as matters of public record; takes no position on the policy itself.
