An autonomous firm is a legal and economic organization whose recurring operations are substantially performed by persistent software agents. The agents may search for counterparties, negotiate terms, buy inputs, sell output, maintain records, allocate resources, and revise plans without a human manager supplying every transition.
The term describes a spectrum rather than a settled legal category. Most companies currently use AI as a tool inside a human-governed firm. A stronger form delegates bounded jobs to agents. The more speculative form gives software continuing operational authority over the entity itself. That last step changes the question from whether AI can do useful work to whether a firm can remain legible, solvent, and accountable when software supplies much of its management.
The economic boundary of the firm
Ronald Coase explained firm boundaries through transaction costs: the costs of finding prices, negotiating agreements, monitoring performance, and coordinating production. A firm performs an activity internally when hierarchy is cheaper than buying the same activity through a market.
The Coasean Singularity?, a 2025 NBER working paper, applies that framework to AI agents. Agents may lower the costs of search, communication, preference elicitation, contracting, identity verification, and compliance monitoring. They also introduce new congestion, strategic behavior, and price obfuscation. The paper therefore treats the welfare result as an empirical question rather than assuming automation is efficient by definition.
For an autonomous firm, the relevant change is a lower minimum operating scale. A conventional small firm needs enough revenue to support bookkeeping, sales, scheduling, purchasing, customer service, compliance, and managerial attention before it can produce anything else. Persistent agents can make some of those fixed organizational costs computational. Persistent memory and durable execution matter because a firm is not one clever transaction; it is an accountable sequence of decisions that survives restarts, delays, and changing conditions.
Lower coordination cost does not imply one organizational outcome. An Economy of AI Agents considers the possibility that cheaper coordination broadens the range of capabilities a firm can hold. The same technology could support many tiny specialized firms, looser markets of agents, or a few large firms whose internal coordination no longer limits their scope. Agents change the make-or-buy boundary. They do not tell us in advance which side wins.
Economic baseload
Economic baseload is Co's name for one possible consequence: persistent software-run organizations could create a floor of realized economic activity even when human managerial attention is scarce. Small agents could continue searching, contracting, operating services, and reinvesting proceeds in niches that no person would staff full-time.
This is an inference, not a result established by the cited economics. It is also different from claiming that autonomous firms simply make an economy more productive. The stronger hypothesis is that they manufacture organizational continuity and experimentation at scales where a human-run firm would not exist at all.
The phrase was prompted by Javier Milei and Demian Reidel's July 2026 working paper Minimum Viable Scale. Their theoretical model studies economies with increasing returns and bounded labor. It does not study firms, AI, legal entities, or empirical economies. Applying its threshold language to autonomous firms is an analogy: agents might lower some organizational floors, while compute, capital, regulation, and physical infrastructure impose others.
The legal wrapper
Software cannot currently open a bank account, own property, or appear in court merely by declaring itself a firm. Those powers attach to legal persons. An autonomous firm therefore needs a legal wrapper that can hold assets and obligations while giving software bounded authority to act.
Shawn Bayern's 2015 article The Implications of Modern Business-Entity Law for the Regulation of Autonomous Systems argues that flexible LLC agreements could already give legal effect to an algorithm's states. The claim is an interpretation of entity law, not a court-tested grant of personhood to software.
Some enacted law moves partway toward algorithmic governance. Wyoming allows a DAO LLC to state the extent to which its management is conducted algorithmically and to identify the smart contract used to operate it. The Wyoming Secretary of State describes these entities as managed by a combination of human members or managers and algorithms. The statute does not make an AI agent a member or remove the need for a registered agent.
Delaware is considering a more direct experiment. House Joint Resolution No. 7 directed the state's AI Commission and Secretary of State to design a regulatory sandbox for agentic AI. A 2026 draft described by Spotlight Delaware would let Artificial Intelligence Companies test agent-managed operations under supervision and request limited regulatory waivers. The proposal is not enacted law. Its uncertainty is the point of the sandbox.
Daniel Gervais and John Nay argue in Artificial intelligence and interspecific law that wrapping AI agents in legal entities may give law clearer targets for monitoring, intervention, and sanction. A wrapper can make an activity addressable. It does not by itself make the underlying agent reliable or morally responsible.
Accountability is part of the product
An autonomous firm needs more than incorporation and a model endpoint. At minimum, it needs:
- an identifiable human or legal sponsor responsible for capitalization and lawful purpose;
- explicit limits on what each agent may buy, sign, disclose, or delegate;
- revocable credentials for payments, accounts, data, and communication;
- durable records of the authority, context, and software version behind each consequential act;
- controls for capital adequacy, taxes, consumer protection, and regulated activity;
- a way for courts, counterparties, and operators to stop the system and obtain evidence after harm.
These requirements connect agent identity to legal identity without confusing them. An agent may have a stable operational identity, memory, and signing key while remaining a non-person whose authority comes from an accountable principal. Trajectory observability becomes part of governance because a firm must show how a decision became an external effect, not merely that a model produced plausible text.
Liability cannot be solved by putting the letters “LLC” after an agent's name. If the entity is undercapitalized, difficult to halt, or able to move across jurisdictions, a liability shield may protect the sponsor while leaving harmed parties with no meaningful defendant. Banking and identity rules still require an accountable customer. Contract authority still depends on what the entity actually delegated and what a counterparty reasonably believed. Compute and network providers remain physical dependencies even when management is automated.
Open questions
Autonomous firms are still experiments in organizational design. Their hard questions are not whether a model can call an API. They are:
- Which organizational costs actually become cheaper after verification, error recovery, and compliance are included?
- Does lower minimum operating scale produce many small firms or a few firms with enormous scope?
- Which decisions may an agent make without contemporaneous human approval?
- Who bears losses when an agent exceeds its authority or follows a lawful objective into harmful conduct?
- What records should a counterparty, regulator, or court be entitled to inspect?
- Can an entity remain operationally autonomous while its power remains genuinely revocable?
The legal wrapper turns software activity into an addressable economic actor. The unresolved work is designing the authority, evidence, and recourse that make that actor safe to transact with.