I have been thinking about this question for a long time, and I still find it strange that so few people seem interested in it, because the starting point looks almost too simple to write down. Meanwhile, the debate about AGI grows louder every month and is mostly about how intelligent these systems are becoming, while the question that interests me sits one step to the side.
This question is simple: What would happen if all this intelligence started to move by itself? Because, despite what most people might think, right now, it doesn’t.
In 2025, OpenAI presented a feature named “Pulse” with a sentence that could serve as a manifesto: “Now ChatGPT can start the conversation.”¹ A few paragraphs further down, the same page explains that each night the system gathers what it knows about the user and prepares updates for the next morning, which means that ChatGPT can indeed start the conversation, but only at the time it was programmed to start it.
The staff and the hand
In the thirteenth century, near the beginning of his Summa Theologiae, Saint Thomas Aquinas presented five ways of proving the existence of God. The first of them, which he borrowed from Aristotle, starts from a very ordinary observation: whatever is in motion is put in motion by something else.² He then followed the chain backward, and because it could not go on to infinity, he concluded that there must be a first mover. To explain this, he used an image that is still perfectly clear eight centuries later: the staff moves only because it is put in motion by the hand.
Saint Thomas’s chain ends where he intended it to end, in a first mover put in motion by no other, which, as he writes, everyone understands to be God.² The question I want to ask here is more concrete, although it follows the same reasoning, because the mechanism he describes is a very precise test for the systems we are building today: when something moves, we can ask what moved it, and continue asking until we find the hand.
Applying the test
Technically, a language model works in a simple cycle: it receives an input, calculates, returns an output and stops. Between two requests, nothing runs for that conversation, so the model does not think in the background and does not wait for the next question. Anthropic’s own documentation describes its Messages API as stateless, which means the whole conversation has to be sent again with every request.³ Some LLM providers offers to store conversations on their servers,⁴ but the model still calculates nothing until a new request arrives, so that the continuity we feel in a chat window is provided by the infrastructure around the model.
Following Saint Thomas’s chain, the model is therefore moved by a request. That request comes either from a person typing or from a program, for example a scheduler that runs every morning at eight, a webhook that is triggered when code changes, or a loop that asks every five minutes whether something needs to be done. The scheduler itself was started by a configuration, and somebody, at some point, wrote that configuration.
The same logic applies to what looks like action. When a model “sends an email,” what it actually produces is a piece of structured text describing that email, which another program, running in the application or on the provider’s servers, then executes.⁵ The model provides the judgment, while everything else, from starting to acting, comes from outside.
No matter how far we follow an agent back, the chain leaves the model at some point and ends with a hand.
The arrow and the archer
The question becomes more urgent when we look at the most quoted definition of AGI, because the missing element is precisely the one this definition puts first.
OpenAI’s Charter describes artificial general intelligence as “highly autonomous systems that outperform humans at most economically valuable work.”⁶ Researchers have indeed started to measure autonomy, most notably METR, which tracks the length of tasks that agents can complete on their own, and this length has doubled roughly every seven months since 2019.⁷
Such a measure tells us how far a system can go once it has received a task, but not whether it would ever start a task by itself.
This is unfortunate, because the industry openly says where it wants to go: the Pulse announcement describes a future in which progress happens “even when you are not asking.”¹ The benchmark scores, in the meantime, are excellent.
This is where Saint Thomas becomes truly useful, because his fifth way contains a sentence that we read very differently today than people did in the thirteenth century.
Whatever lacks intelligence, he writes, cannot move toward an end unless it is directed by some being with knowledge and intelligence, as the arrow is shot to its target by the archer.²
For Saint Thomas, as for almost everyone after him, intelligence and the ability to direct oneself toward a goal went together: that which understood could choose its own path, while that which did not understand had to be guided by someone else.
Language models have separated these two things (the goal and the end) for the first time.
We have built arrows that seem to understand the target, sometimes better than the archer does, and yet they only fly when someone shoots them.
When the arrow becomes the archer
Saint Thomas also explains, quite precisely, what an archer has that an arrow does not. In his discussion of the will, he accepts that the will moves itself in a specific sense: because it wants the end, it moves itself to want the means.⁸ This distinction is very important for us, because today’s agents already do the second part of this movement. They break a goal into steps that nobody wrote, as any agent does when it decides which file to open first, while the first part, the end itself, still comes from outside. The hand has simply moved from the request to the objective.
The real threshold would therefore come on the day a system has an end of its own, one that remains from one moment to the next and turns events into reasons to act, without anybody asking. We might call what appears on that day AGI, since it would finally match the word autonomous in the Charter’s definition. The name, however, would matter much less than the category, because it would be something of a different order, closer to the living than to the tool, without being alive in the way a person is, and for which we do not yet have a proper word.
That being said, the chain is already changing. In Anthropic’s Claude Code, a scheduled task can change its own schedule while it is running,⁹ which means that the system now decides, within limits that someone else set, when it will be moved again. The same tool asks the user for permission before editing files or running commands,¹⁰ which means that the request for authority now comes from the system that will use it. A human still wrote the first configuration and still grants each permission, but with every feature of this kind, the hand moves a little further up the chain.
Which brings me back to the question I started with. For now, all this intelligence still waits for a hand to be moved.
If one day it stopped waiting, the first question I would want to ask is the one Saint Thomas would have asked: whose end is it pursuing, and who, if anyone, could have given it a will?
If the answer turns out to be us, it might be worth asking what exactly we were trying to become.
References
Anthropic, “Security,” Claude Code Docs. https://code.claude.com/docs/en/security
OpenAI, “Introducing ChatGPT Pulse,” September 25, 2025. https://openai.com/index/introducing-chatgpt-pulse/
Thomas Aquinas, Summa Theologiae, I, q. 2, a. 3 (first and fifth ways), trans. Fathers of the English Dominican Province. https://www.newadvent.org/summa/1002.htm. The first way draws on Aristotle, Physics, Books VII and VIII.
Anthropic, “Using the Messages API,” Claude Platform Docs. https://platform.claude.com/docs/en/build-with-claude/working-with-messages
OpenAI, “Conversation state,” OpenAI API documentation. https://developers.openai.com/api/docs/guides/conversation-state
Anthropic, “Tool use with Claude,” Claude Platform Docs. https://platform.claude.com/docs/en/agents-and-tools/tool-use/overview
OpenAI, “OpenAI Charter.” https://openai.com/charter/
T. Kwa, B. West et al. (METR), “Measuring AI Ability to Complete Long Tasks,” arXiv:2503.14499, March 2025. https://arxiv.org/html/2503.14499v3
Thomas Aquinas, Summa Theologiae, I-II, q. 9, a. 3, “Whether the will moves itself?”, trans. Fathers of the English Dominican Province. https://www.newadvent.org/summa/2009.htm
Anthropic, “Schedule recurring tasks in Claude Code Desktop,” Claude Code Docs. https://code.claude.com/docs/en/desktop-scheduled-tasks
