In July 1935, nearly five decades after organized labor emerged as a coherent economic force in the United States, Congress passed the National Labor Relations Act.[1] The NLRA established the legal framework for collective bargaining, union elections, and worker protections that labor advocates had sought since the 1880s. The delay was not accidental. It reflected something consistent across modern economic history: New actor types tend to operate in a governance vacuum for a generation or more before the institutions that constrain and legitimize them finally arrive. Organized labor, the modern corporation, and the independent central bank all followed this arc. AI agents, as autonomous economic and software actors, appear to be following it now.
New Actors, Governance Vacuums, and Eventual Frameworks
The structure of governance transitions for new economic actor types is surprisingly consistent across the modern period. An actor type emerges, typically enabled by new technology or legal innovation, and begins to accumulate economic significance. Existing institutions are not equipped to govern it. The actor operates in a governance vacuum, sometimes for decades. Something goes badly wrong. And then, eventually, the frameworks arrive.
Organized labor offers the clearest case. The American Federation of Labor was founded in December 1886, the same month as some of the most violent confrontations between workers and state power in American history.[2] The Haymarket affair, in which a bomb thrown during a labor rally in Chicago killed eight police officers and four workers, brought the question of labor organizing into sharp public relief.[3] The Pullman Strike of 1894 required federal troops and the imprisonment of union leader Eugene Debs before it ended.[4] The Triangle Shirtwaist Factory fire of 1911, which killed 146 workers in a building with locked exit doors, made the human cost of unregulated industrial labor undeniable.[5] The NLRA arrived in 1935, roughly fifty years after the AFL’s founding, establishing the protections that had been contested through violence for generations.
Corporate power followed a similar arc. The modern corporation as a legal entity capable of accumulating capital and operating across jurisdictions developed through the 19th century, and the robber baron era demonstrated both its economic power and its social costs. The Sherman Antitrust Act arrived in 1890, the first major federal attempt to constrain that power.[6] The Clayton Act followed in 1914, adding substance to Sherman’s broad prohibitions.[7] The Securities Exchange Act of 1934 established the oversight frameworks for capital markets that unconstrained corporate power had made necessary.[8] From the rise of the modern industrial corporation to basic regulatory infrastructure: Roughly half a century.
Central banking is perhaps the most instructive example, because it involves the explicit question of institutional independence. That is, whether a technical actor can be insulated from short-term political pressure in order to serve a long-term public function. The Federal Reserve was established by Congress in 1913.[9] But its operational independence from the Treasury Department was not formalized until the Treasury-Federal Reserve Accord of March 1951, a document that established the basic principle that monetary policy should be free from Treasury influence.[10] Even after the Accord, the normative consensus that central banks should be genuinely independent actors did not fully crystallize until Paul Volcker’s disinflation of the early 1980s. The governance framework for what kind of actor a central bank should be took the better part of seven decades to settle.
AI Agents Are in the Governance Vacuum Phase
Autonomous AI agents, as actors that write code, execute financial transactions, manage infrastructure, and increasingly make consequential decisions without real-time human oversight, are currently in the governance vacuum phase.
The frameworks for governing this class of actor do not yet exist in any meaningful form. There is no settled framework for liability when an AI-generated smart contract contains an exploit. There are no audit standards for AI-authored code deployed in critical systems. There is no clear accountability chain when an autonomous agent fails in a way that causes financial or physical harm. The EU AI Act, which entered into force in August 2024, establishes risk categories and some baseline requirements for high-risk AI systems, but it does not address the specific accountability questions raised by autonomous agents acting without direct human direction.[11] The Biden Administration’s Executive Order 14110, signed in October 2023, directed federal agencies to develop standards and guidance but created no binding accountability frameworks for autonomous agents operating in the private sector.[12] The G7 Hiroshima AI Process, launched in May 2023, produced voluntary principles with no enforcement mechanisms.[13]
These are early signals of governance attention, not governance frameworks. The gap between signal and framework, measured against historical precedent, is likely to be long.
What Is Different This Time
Three factors complicate the historical analogy and deserve serious attention before assuming the 50-year arc will simply repeat itself.
Speed
Corporate personhood took centuries to develop as a legal concept. The organized labor movement took decades to coalesce into a political force capable of demanding legislative attention. AI agents went from a research curiosity to a deployed economic actor in roughly five years. This compression of the development timeline may shorten the governance vacuum. It may also accelerate the arrival of the precipitating crisis. Whether that produces better or worse governance outcomes depends largely on whether the intervening period is used well.
Technical Opacity
When a bank fails, regulators can trace the balance sheet. When a labor dispute breaks out, arbitrators can read the contract. When a corporation engages in anticompetitive behavior, antitrust enforcers can reconstruct the business decision. When an autonomous AI agent makes a consequential and harmful decision, the chain of accountability is genuinely unclear. Current frontier AI systems make decisions through processes that are not fully interpretable even by the teams that built them. This is not a solvable engineering problem in the near term, and it creates a fundamental challenge for governance frameworks built on the assumption that accountability requires legibility. Governance for AI will need to grapple with a form of opacity that has no good historical precedent.
Cross-Jurisdictional Complexity by Default
Corporate law is jurisdictional. Labor law is national. Central banking is sovereign. These forms of governance worked, imperfectly and with significant friction, because they could be applied within the territorial boundaries of the state. AI systems operate across jurisdictions as a design characteristic, not as an exception or an evasion. A model trained in one jurisdiction, deployed through infrastructure in a second, used by operators in a third, and affecting individuals in a fourth does not map cleanly onto any existing national or regional governance framework. The EU AI Act applies only within the EU. Executive orders apply only to US federal agencies and contractors. The G7 Hiroshima Process produces only voluntary principles. No existing governance instrument addresses the cross-jurisdictional nature of AI as a systemic matter, and the political conditions required to create one do not currently exist.
Reactive Is the Default, and the Opportunity Is to Do Better
The Sherman Antitrust Act was a response to Gilded Age monopolies and their visible social costs. The Wagner Act was a response to decades of labor conflict and a political moment in which the costs of the status quo were undeniable. The Volcker reforms and the consolidation of central bank independence were a response to stagflation, a concrete demonstration that political interference in monetary policy produced bad outcomes. In each case, the precipitating event created the political conditions for governance frameworks that had been technically possible for years but politically infeasible until the crisis made them necessary.
For AI agents, the precipitating event has not yet happened. The governance vacuum will persist until it does. My expectation is that it will take the form of a significant incident involving autonomous agents making consequential financial or legal decisions at scale, and that the incident will precipitate serious regulatory attention in short order. What the frameworks that emerge from that moment look like will depend significantly on the quality of thinking done in advance.
There is one genuine difference between AI and its historical analogies that carries at least some optimism. The people building the systems that will eventually precipitate the governance crisis are, in many cases, also actively engaged in thinking about what responsible governance would look like. That has rarely been true historically. The captains of industry who built the monopolies of the Gilded Age were not collaborating with antitrust scholars. The factory owners of the early 20th century were not designing labor law. The financiers of the pre-SEC era were not writing securities regulation. The technologists building AI systems are, with imperfect and uneven results, at least in the room when governance questions are discussed.
Whether that engagement produces better governance frameworks, or is simply absorbed as a delay mechanism, will say a great deal about whether this generation of technologists learned anything from the governance failures that preceded them.