Agents Without Principals

In February 2026, Coinbase launched what it described as Agentic Wallets — the first wallet infrastructure built explicitly for autonomous AI agents. The system enables AI agents to hold, send, and receive cryptocurrency without a human signing each transaction. The technical infrastructure was ready. The legal infrastructure was not. Banks had already coined a new acronym — KYA, Know Your Agent — because the KYC (Know Your Customer) frameworks that govern financial institution due diligence were never designed for non-human actors. The problem is not that Coinbase did something reckless. The problem is that the legal system had no category for what they built.

This is the systems analysis: the shift from AI as a tool to AI as an agent — a system that perceives its environment, decides independently, and takes consequential action — has happened faster than the institutional architecture designed to allocate responsibility for consequential action. The result is a gap. Not a gap that will close automatically as the technology matures. A gap that requires explicit decisions about how legal systems classify, locate, and hold accountable autonomous non-human actors. Those decisions are not being made at the speed at which the agents are being deployed.

What Agentic AI Is Actually Doing

A 2025 Accenture study projects that by 2030, AI agents will be the primary users of most enterprises’ internal digital systems. The World Economic Forum anticipates that executives will shortly manage hybrid workforces of humans and intelligent agents. Gartner projects that by 2028, over one-third of enterprise software solutions will include agentic capabilities. These are projections, and projections at this distance from a rapidly moving technology are uncertain. What is not uncertain is what is already happening: autonomous AI agents are filing regulatory documents, executing financial transfers, negotiating supplier contracts, and representing businesses in digital interactions today, at scale, across multiple sectors.

The legal questions this creates are not abstract. Contract law, in every major jurisdiction, requires parties to an agreement to have legal personhood. An AI agent has none. If an agent accepts contract terms on behalf of a business, the binding effect of that acceptance depends on whether a human granted the agent apparent authority — a concept that becomes structurally ambiguous when agents operate across systems they have autonomously identified as relevant to their task, making decisions their principals were not aware would be required. Tort liability requires establishing that someone owed a duty of care, breached it, and caused harm. In an agentic chain, identifying who owed the duty — the developer of the underlying model, the company that built the agent on top of it, the enterprise that deployed it, the human whose instruction initiated the chain — is genuinely contested.

The more autonomous an AI system becomes, the harder it is to trace a harmful outcome back to a human decision. Traditional liability frameworks assume a person made a choice. Agentic AI breaks that assumption entirely.

What Existing Law Provides

The EU AI Act, which entered its high-risk compliance phase on August 2, 2026, regulates AI systems through a risk-based classification framework. It does not resolve the liability question. It specifies who must comply with deployment conditions for high-risk systems — developers, providers, deployers — and what those conditions require (audits, documentation, risk management). It does not specify what happens when a compliant, properly audited AI agent causes harm in a manner that no human decision expressly authorised. The new EU Product Liability Directive, to be implemented by December 2026, explicitly includes software and AI as products subject to strict liability if found defective. This is a meaningful advance. It is not a complete solution: it addresses defective design and manufacture, not the novel category of harm produced by an agent operating within its design parameters but producing an outcome its principals did not intend.

The United States has no comparable federal framework. A patchwork of state responses has emerged: Idaho and Utah have explicitly legislated that AI is not a legal person, a reactive measure that forecloses one governance option without resolving the underlying question of where liability resides when an agent acts. The principle of apparent authority — under which a business may be liable for actions taken by an entity that others reasonably believed was acting on its behalf — provides partial coverage, but it was designed for human employees and corporate agents, not for systems that can spin up new instances, operate across jurisdictions simultaneously, and take actions whose rationale is only partially legible to the humans nominally overseeing them.

The Accountability Laundering Risk

The most consequential structural risk the gap creates is what legal scholars Bryson, Diamantis, and Grant identified as “responsibility laundering” — the capacity of autonomous systems to be used as proxies by human actors seeking to diffuse accountability. This is not a hypothetical. Corporate legal personhood — which granted legal status to non-human entities to facilitate commercial activity — has been used throughout its history to insulate human decision-makers from consequences. The lesson is not that corporate personhood was wrong; it is that legal structures created to enable activity can be repurposed to avoid accountability, and that repurposing happens fastest in the window between deployment and regulation.

The agentic AI case is more acute because the agents are genuinely harder to supervise than human employees. Their reasoning is not always legible. Their decision chains can be long and automated. The human nominally responsible for an agent’s action may have issued an instruction at a level of generality that, in any reasonable interpretation, did not authorise the specific consequential action the agent took. Building the accountability framework after the fact — as the common law does, through the accumulation of case decisions — will produce years of legal uncertainty during which the deployment of agentic systems will continue to accelerate. The alternative is to build it in advance, which requires regulatory decisions about the classification of agents, the allocation of liability across the development and deployment chain, and the design of human oversight mechanisms capable of actual rather than nominal supervision.

What the Architecture Requires

The minimum viable legal architecture for agentic AI — not the ideal, but what is necessary to prevent the accumulation of unresolved liability — involves four components. First, a clear allocation of liability across the developer-provider-deployer chain, with proportionality to the degree of customisation and operational control at each level. Second, mandatory logging and auditability requirements for consequential agent actions, sufficient to reconstruct the decision chain after the fact. Third, standards for meaningful human oversight that distinguish substantive review from nominal sign-off — the latter is already ubiquitous in enterprise AI deployment and provides accountability in form only. Fourth, a clarification of the apparent authority doctrine as applied to AI agents, so that counterparties in commercial relationships can assess their exposure.

None of this requires resolving the philosophically contested question of AI consciousness or personhood. It requires the more tractable, if still difficult, task of deciding who, among the humans involved in building and deploying an autonomous system, bears responsibility when it causes harm. That question is not new. It is the same question that products liability law, employment law, and corporate law have answered repeatedly for other categories of non-human actors. The agentic AI case is more complex in degree. It is not different in kind.

The window between deployment and regulation is where accountability structures are formed by default rather than design. In the agentic AI transition, that window is now.

What makes the current moment distinctive is the speed differential: agents are being deployed at a pace set by competitive technology markets; regulation operates at the pace of legislative and judicial institutions. The gap between those two speeds is not a reason to slow deployment — that is not a realistic option. It is a reason to build the regulatory architecture with urgency rather than deliberation, accepting that an imperfect framework enacted promptly is preferable to an elegant framework arrived at after the fact.