Agency - in the AI age

Human agency
in the AI age

You ask an assistant which of two offers to send a client. It recommends the second in four tidy paragraphs, and you send it. A week later you cannot say which part of the reasoning you checked.

The scene opens the technology chapter of Modern Hermet. The book says it is a composite, not a report. We usually put it down to laziness. The chapter sees an asymmetry: a system built to sound sure, facing a well-documented human tendency to accept a confident prompt. That is where the question of agency starts: how to hand a model the work and still be the one who answers for it.

What AI can take, and what stays with you

The chapter splits the work into three layers and lets only the first two through.

• Execution. “Delegate without remorse.” Rewriting, ordering, first drafts, summaries of long documents, the scaffolding of code, ten variants you will throw away.
• Judgement. “Delegate the input, never the output.” Ask the model for arguments against your thesis, for other framings, for what you are not seeing. Which framing is true stays on your side, “because the model does not answer for consequences. You do.”
• Intention. “Here delegation ends.” What you want, what is worth it, whether the project should exist at all.

The chapter separates attention from judgement in one sentence: “the capture of attention is uncomfortable, the capture of judgement is another category.” Its test is a question you can check: whose intention is this tool carrying out just now? Somebody’s criteria sit inside every system, in the success metric, in the default setting, in what is one click away and what is seven. The chapter also points to research on automation bias, Parasuraman and Riley (1997) and Skitka and colleagues (1999): people using automated prompts miss what the system did not flag, and carry out what it flagged wrongly. In the chapter’s words, trust in a tool grows faster than the tool’s accuracy.

The chapter on agency says where to push. Agency “is not willpower”; it is finding the place where a small input produces a disproportionate change. With AI that place is usually dull: “ask what you do by hand every week that has a clear criterion of correctness.” A tool that only gives you more to review has not removed the work, only moved it.

Four moves for this week

Each takes minutes. The arrangement is ours, drawn from the book and from the Seven Moves. Each move says why it should help and where that stops.

1. Write three sentences before you ask. Before you open the model, write what you came for and what you think now. The book’s rule: do not give away first contact with a problem. You hold a position, so the answer becomes material rather than revelation. Writing down why you came also gives you something to check the answer against. No study cited here tests this exercise.
2. Separate your state from the facts. For whatever is nagging you, three lines before you hand it over: my state, the facts, what I am adding to them. Separating the feeling from the facts lowers how much the feeling steers the reading of the facts. The same separation is part of cognitive therapy, whose efficacy as a treatment is well documented (Hofmann et al. 2012), and in the laboratory naming a feeling measurably dampens the response to it (Lieberman et al. 2007). Three lines are not therapy: those studies make the mechanism plausible, they do not show that this exercise works. The Pre-AI Orientation file does the same in five questions.
3. Write down what AI does not do for you. This is day two of the book’s seven-day AI charter. Two of its examples: it does not write messages to people you care about, and it does not take decisions whose consequences you will bear. Add one threshold concrete enough to check, like the book’s “the last twenty per cent of anything under my name I do by hand”, which it offers as a working rule rather than an optimum. A boundary written in advance spares you deciding it again twenty times a day, each time in a worse state.
4. End each day on one word, and review the week. Author or material: which were you today? On day seven ask the charter’s exit question: did the tool strengthen your agency this week, or take it over? Correct one line of the charter. Habits take weeks to months to settle (Lally et al. 2010), so judge the practice on that timescale, not on one good or bad day.

Where this comes from

The question is older than any machine: how much of a life does a person steer? One of the oldest and most direct answers in the West comes from the Hermetica, Greek texts written in Roman Egypt around 100 to 300 CE. Source, reading, practice and evidence are different things, so they are kept apart here.

• Source. In treatise IV God sends down a mixing-bowl of nous, mind; those who choose to be immersed in it gain mind, the rest remain merely alive. The Asclepius, which survives complete only in Latin, calls the human being a great miracle, a creature that joins the divine and the material. The same text describes men drawing power into statues so that the object begins to act. In 1486 Pico della Mirandola opened his Oration with the great-miracle line; the speech was never delivered.
• Our reading. In its own setting the statues passage is about ritual. Modern Hermet reads it, on its own account, as an early image of an artefact with an intention built into it, and adds: “That is precisely software.” That reading is ours, not the text’s. We call Hermeticism the oldest school of the agency question because of treatise IV’s choice to take mind or leave it, not because of the statues. It is not a school of technology.
• Practice. The texts hand out no programme. Poimandres reports a vision and gives no steps. The four moves and the AI charter are our arrangement.
• Evidence. The studies cited above tested mechanisms, one of them as a clinical treatment, not these moves. They support the everyday psychology, not the Hermetic framing.

What this does not settle

Whether the moves will work for you. They have not been tested as a set. The book names two limits of its own charter: it does not work against tools an employer requires, which are a matter for a conversation or a decision about the job, and it does not replace competence. No written rule protects you from accepting a wrong answer in a field you do not know well enough to spot the error.

Nor does it settle what happens to work in general. The chapter’s bet, that intelligence becomes cheap while taste, intention and judgement grow expensive, is in its own words “a forecast, not a finding.”

Where to go next

The whole argument is in the open chapter, Technology, free in full. The rest is in the book: Modern Hermet, twelve chapters with one practice each. The method in short form is the Seven Moves; for goals, how to manifest, technically. And the letter, As Above: one source, one correction, one map and one question at a time. Get the letter.

Fair questions

What is human agency?

Your capacity to set the direction of your own work and life and to act on it. Modern Hermet gives a narrower working definition: agency is not willpower but the recognition of the place where a small input produces a disproportionate change of state, and the readiness to bear that input.

How do I keep my judgement when I use AI?

Delegate the input to your judgement, never the output: ask the model for counter-arguments and other framings, then decide yourself. Write three sentences of your own before you ask, and at the end of the day note whether you were the author or the material.

What should I not hand over to AI?

Intention: what you want, what is worth doing, whether a project should exist. And messages where your presence is the point, such as condolences, an apology or feedback for an employee. Generated, they stop being what they were meant to be.

Is this an argument against AI?

No. The book’s chapter on technology says the modern hermet “is not anti-technological. He is anti-capture.” Its closing advice is to use a great deal and choose narrowly: give the machine everything that is execution and nothing that is choice.

Sources & further reading

• Modern Hermet (The Hermeticum, 2026), chapter 8, Technology (open in full), and chapter 11, Agency. Quotations on this page come from these two chapters.
• Parasuraman and Riley (1997) and Skitka and colleagues (1999) on automation bias, as cited in chapter 8.
• Hofmann, S. G. et al. (2012). The efficacy of cognitive behavioral therapy: a review of meta-analyses. Cognitive Therapy and Research 36(5), 427-440. doi:10.1007/s10608-012-9476-1
• Lieberman, M. D. et al. (2007). Putting feelings into words. Psychological Science 18(5), 421-428. doi:10.1111/j.1467-9280.2007.01916.x
• Lally, P. et al. (2010). How are habits formed: modelling habit formation in the real world. European Journal of Social Psychology 40(6), 998-1009. doi:10.1002/ejsp.674
• Brian P. Copenhaver, Hermetica: The Greek Corpus Hermeticum and the Latin Asclepius (Cambridge University Press, 1992): treatise IV, Poimandres and the Asclepius.
• The four moves are our own arrangement. They are not a treatment and make no therapeutic claim.