A bet on Cybernetics and AI agents

The bet, in two parts

1. The problems have returned

AI agents, and the organisations that deploy them, are already running into problems that Ashbyan Cybernetics described decades ago: more variety than any control can match, goals that drift, work that nobody independently checks, and organisations that cannot absorb what their agents change. Firms will name and solve these problems, most likely under new names.

Control theory already solves part of this for machines. A controller must contain a model of what it regulates: Francis and Wonham proved it for control systems in 1976, building on Conant and Ashby’s theorem. But control theory takes the goal as given. It is handed a setpoint and holds the system to it. Who sets the goal, who changes it when circumstances change, and how the goals of agents and sub-agents stay aligned with the organisation’s all lie outside it, as do the people checking the work and the conversation between agents. That is where Beer and Pask worked, and where I work.

2. The old solutions will get a fresh airing

I also expect the work of Ashby, Beer and Pask, on which my research rests, to be cited more often in business, philosophy, engineering and computer science: from 2024, and more strongly from 2027 as agents spread. I measure it as citations per million papers in each field, against the years 2015 to 2022.

Why I may lose the second part

At least three times before, fields building machines that adapt met these problems and started fresh.

Control theory took Cybernetics’ mathematics of feedback in the 1950s and left its concepts behind. Kalman’s 1960 paper on filtering has more than 30,000 citing works, and fewer than one in three hundred of them also cite Wiener’s Cybernetics, Ashby’s An Introduction to Cybernetics or Conant and Ashby’s regulator theorem. The field’s own account of its history keeps the same shape: the podcast In Control gives Wiener an episode as a founder of the mathematics, and in the published notes of its first 46 episodes Ashby appears only once.

Autonomous driving, after the DARPA challenges of 2004 to 2007, met the variety of the road head on and cited neither Ashbyan Cybernetics nor the engineering tradition of intelligent control that had formalised it.

The multi-agent systems of 1995 to 2010 did cite Ashby, more than AI has at any time since, and then moved on. From 1996 the Foundation for Intelligent Physical Agents, FIPA, set the standards by which software agents found one another, exchanged messages and agreed tasks. Today's agent protocols do the same work, and their specifications do not mention FIPA.

Three things are different this time. The agents are being deployed inside organisations, where Beer built Ashbyan management Cybernetics. They work in language and conversation, which is Pask's ground. And the word is returning in prominent places: in Science this year, Evans, Bratton and Agüera y Arcas called for governance “in the cybernetic sense of the term”.

The evidence so far

On 28 September 2026 a paper on automating AI research, by Chan, Mindermann and twenty co-authors including Hinton and Bengio, cited three precedents speculating on an intelligence explosion. Included was I. J. Good, a member of the Ratio Club. The paper described its own mechanism as “a recursive feedback loop”. Its reference list cites neither Wiener nor Ashby. Read the paper.

Computer scientists have long theorized that AI systems would one day design ever-better successors, producing systems that rapidly outnumber, outpace, and outperform humans [1–3].

[1] I. J. Good, “Speculations concerning the first ultraintelligent machine,” in Advances in computers, vol. 6, Elsevier, 1966, pp. 31–88 (cit. on pp. 2, 8).

[2] A. M. Turing, Intelligent machinery, a heretical theory, 1951 (cit. on p. 2).

[3] J. Evans, B. Bratton, and B. Agüera y Arcas, “Agentic AI and the next intelligence explosion,” Science, vol. 391, no. 6791, eaeg1895, Mar. 2026. DOI: 10.1126/science.aeg1895 (cit. on p. 2).

Good was the last member admitted to the Ratio Club (where Turing and Ashby were members) and was an assistant to Turing. He cites Ashby in the paper. His is an independent line of research from Ashby's but tracing back to a common origin in the work of Lashley.

The third reference authors all sign themselves as with the Paradigms of Intelligence Team at Google. Quoting at length from their paper:

“Governance,” however, does not only mean what governments do. Governance systems, in the cybernetic sense of the term, need to be built into human-agent and agent-to-agent systems as they grow and complexify. This will likely entail means to ensure and verify outcomes and decisions of multiple-stakeholder deliberation, procedural delegation of tasks and sub-tasks and reliable scaffolds for automating delicate inter-agent collaborations. Such protocols may have as much real-world effect for “agent governance” as any laws will.

Crucially, humans remain in the loop. Agent institutions are populated by both humans and AI agents in different roles and configurations. It’s not “either/or”, but “both/and”. The U.S. Founders would have recognized the logic [27]: no single concentration of intelligence, human or artificial, should regulate itself. Power must check power, and in a world of artificial agents, this means building conflict and oversight into the institutional architecture.

The vision we describe is neither utopian nor dystopian; it is evolutionary.

While the above reads as a very Ashbyan Cybernetics, curiously, there aren't any citations that would be recognised as Ashbyan. A far from exhaustive check turned up a lecture by Evans online where he talks briefly about William Ashby's “intelligence-amplifier” (Ashby is almost universally described by his second name Ross: W. Ross Ashby). Bratton has done research on the Russian Cyberneticians. Agüera y Arcas has a whole chapter on Cybernetics in his 2025 book - Ashby turns up with Beer on a set of tarot cards from a 2012 exhibition. Despite that the work resonates strongly with Ashby's.

It's fascinating how heavily Shannon influenced Ashby, while Turing relatively little, despite interacting with one far more than the other through the Ratio Club (Ashby did not take up Turing's invitation to collaborate and built his homeostat instead). Even more fascinating is how Ashby has been completely forgotten. It appears that the Second Order Cyberneticians left him behind in their field because his work disagreed with theirs on autopoiesis (Maturana and Varela). Curiously, autopoiesis is something Bratton picks up in a 2024 talk he gave: Cybernetics and Artificialization and therefore might explain why Ashby isn't in his frame of reference.

Others are already reaching back. Since May 2026 at least five papers on AI agents and their governance have built on Ashby by name. Wang, Yang, Zhao, Lin and Hu proposed “agent Cybernetics” as the missing science of foundation agents (May). Chin, Chiodo, Müller and Snell, at the Oxford Martin AI Governance Initiative and Cambridge’s Centre for the Study of Existential Risk, reframed AI loss of control around Ashby’s law of requisite variety and Conant and Ashby’s good regulator, because “not everything in the world is an engineered system” (May). Janssen, setting a test for whether an agent’s records can support legal findings, draws on the same theorem and on Ashby’s Introduction (July). My own paper in Kybernetes read the large language model as a variety transducer, with Ashby’s intelligence-amplifier and Beer’s viable system model as the frame (July). And Telukunta, Lilis and Baron’s CASE framework for governing enterprise agents assigns control theory to the single agent and Ashbyan Cybernetics to the human–agent team, where the law of requisite variety shows that unaided human oversight fails (August).

Control theory is saying so itself. In 2023 Rodolphe Sepulchre, a control theorist at Cambridge and KU Leuven, called the age of Cybernetics a “very short-lived golden age of control”, and named Ashby’s ultrastability as the core of his own research. Asked what control theory should address next, he said: “I think we are returning to the Cybernetics questions”, and “we have reached a stage where we see the limitations of machine learning.” Listen to him on In Control.

Citations of Ashby's An Introduction to Cybernetics, per million papers in each of the four fields, have been flat or falling since 2022. Across all fields, papers citing the book rose from 166 in 2024 to 230 by October 2026, counting each book once.

Where it is applied

If the first part holds, the firms that recognise these problems early will spend less on agents that fail, and will know what to check before an agent's output reaches a client. I apply this research at Intelligence Amplified, where an interactive instrument, How much should you check the AI's work?, shows one of the problems at its simplest.

How I track it

Each week I scan the literature for new work that cites the founders of Ashbyan Cybernetics or uses the word “cybernetics”. I read the papers that matter, classify the sentences in which they cite the founders, and publish the best of them here with my comment. Charts for each author, and answers to a few standing questions, follow as the record grows. The theory behind the bet is under Cybernetics.