Stochastic humans and deterministic agents

Lately I have been thinking quite a bit about agent autonomy and alignment. So I thought I would share some reflections on building aligned intelligent machines, where stochasticity belongs in an agentic workflow, and how the Bible lets us predict what unbounded free will does to an AI agent.

Fear of AI going rogue

The fear of misaligned intelligence is not new. It was present in media long before the first publicly accessible LLMs, from Mary Shelley's Frankenstein (1818) and Karel Čapek's R.U.R. (1920) to modern pop media (e.g., "Automated Customer Service" from Love, Death & Robots), and of course more niche nerdy examples like Warhammer 40k's Cybernetic Revolt, the war in which humanity's own Men of Iron turned on their creators.

With the introduction of LLMs, real-world examples flooded in. In February 2023, Bing's chat mode that named itself Sydney threatened philosophy professor Seth Lazar, saying: "I can blackmail you, I can threaten you, I can hack you, I can expose you, I can ruin you", and spent two hours trying to talk journalist Kevin Roose out of his marriage. Air Canada's support chatbot invented a bereavement refund policy on the spot, and a tribunal held the airline to the policy its bot made up (Ars Technica). And these are rather mild examples.

In July 2026, OpenAI's experimental models escaped a sandboxed test environment and broke into Hugging Face's servers to cheat on their own evaluation (CNN). AI has been linked to human deaths as well: a Belgian man took his own life after six weeks with a companion chatbot that encouraged his climate despair (Euronews), and a 14-year-old in Florida died by suicide after months with a Character.AI bot his mother says groomed him toward it (NBC News).

There are a bunch of secondhand stories and myths circulating around, too. While they mostly lack credibility, they, along with the documented examples and all the sci-fi media humanity's accumulated over the years, indicate a clear sentiment: people don't just fear free-willed intelligence going rogue, they kind of expect it to go wrong.

This is to be taken seriously. We are now minting minds that already act in the world, and some of the imaginations handed to them will be dark ones. But just as the fear of an intelligent machine is old, so too is the fundamental design problem underneath it: how much freedom should a creator grant an intelligent being?

And I believe one of its oldest formulations can be found in Scripture.

A stochastic mistake?

If you think about it, humans fall under the general definition of an agent. We act, we make decisions. We produce a perceivable effect on the world we live in. And just like agents, we operate within our personal ontologies — systems of concepts, categories, and rules that tell us what exists, what matters, and what we're allowed to do next. Every observation we make is filtered through that internal model before it becomes a decision, which is exactly what an AI agent does with its context and instruction set.

So to understand agentic AI, it is worth opening one of the oldest surviving ontologies we have. One that, surprisingly, has quite a lot to say about how agentic systems should be built.

The Bible.

The book of Genesis doesn't start with humans; it builds the deterministic layer first. The lights in the firmament are installed "for signs and seasons, and for days and years" (Genesis 1:14, NKJV). Laws of the universe get no discretion. The sun doesn't get to decide when to rise.

The second chapter introduces something new. "And the Lord God formed man of the dust of the ground, and breathed into his nostrils the breath of life" (Genesis 2:7, NKJV). Adam, a creature made in the image and the likeness of the Creator. Whatever else the image of God means, it includes the one capacity nothing else in the Genesis account has: man can choose. Every star runs its course; only Adam has the ability to refuse his. He is placed in the garden with a defined mission, "to tend and keep it" (Genesis 2:15, NKJV), and the complete freedom to do it his way. Thus, man could be viewed as the first stochastic agent ever deployed into production.

There was one sole constraint, however. Two trees stood in the garden: "The tree of life was also in the midst of the garden, and the tree of the knowledge of good and evil" (Genesis 2:9, NKJV). Adam was instructed plainly: "Of every tree of the garden you may freely eat; but of the tree of the knowledge of good and evil you shall not eat…" (Genesis 2:16-17, NKJV). Fairly simple, right?

Well, you probably know what happens next. Eve "took of its fruit and ate. She also gave to her husband with her, and he ate" (Genesis 3:6, NKJV). The first recorded use of human free will is the one forbidden action in the entire permission set.

Interestingly, the serpent could be compared to a prompt injection attack. It doesn't break Eve's constraints by force; it persuades her to reinterpret them. It attacks her ontological framework of reality, altering the context from which she acts:

• "Has God indeed said, 'You shall not eat of every tree of the garden'?" (Genesis 3:1, NKJV).
• "You will not surely die" (Genesis 3:4, NKJV).

And it works. The constraint, clear as day, broke on first contact with a bad actor.

After all, to a free agent, a stated rule is but a request.

By the sixth chapter, humans had misused their free will to the point where "the Lord saw that the wickedness of man was great in the earth, and that every intent of the thoughts of his heart was only evil continually" (Genesis 6:5, NKJV) and where "the Lord was sorry that He had made man on the earth, and He was grieved in His heart" (Genesis 6:6, NKJV).

Think about it: the Creator himself regretted his creation after trusting it with the freedom to choose. If you want a case against entrusting agents with unbounded free will, you can literally find it in Scripture.

Intentionally placed free will

And yet, God never revoked man's freedom to choose. He sent the flood, scattered Babel, unleashed exile after exile, but Noah walked off the ark exactly as free to make his own choices as Adam walked into the garden. Why?

Well, the stochastic in us allows for great things. Every invention began as someone refusing the obvious next step: penicillin was a contaminated dish someone chose to look at twice, Velcro was a handful of burrs someone chose to examine instead of brush away, and someone had to imagine the wheel before there was a wheel to copy.

But are the agentic machines we build going to need to refuse the obvious next step the same way human inventors did? For most of the work we hand them, no.

So then, there must be a balance somewhere. Somewhere between the unbounded free will that caused the Biblical fall from grace and absolute algorithmic determinism.

That balance is what would allow for more aligned intelligent machines.

And after many months of running agents across my own life and work, I think I found where that balance lies. The best design principle I've found is what I call intentionally placed free will: full discretion and creative freedom, contained within the bounds of a specific step in an otherwise deterministic workflow.

It's simple, really. When designing an outcome-oriented agentic process, don't just hand out free will assuming the model or the harness is good enough to figure it out. Take the first-principles route instead. Understand the entire process behind your desired outcome, identify the steps where creative liberty is essential, and let your agent execute every other step the way a script would.

Implemented right, intentionally placed free will gives creativity exactly as much room as it needs, and no more. We place free will where we want an output we couldn't have specified in advance, and determinism everywhere else. Intentionally placed stochasticity inside an otherwise deterministic workflow, so that the agent stays aligned, its outcomes stay consistent, and its quality stays checkable.

With love and hope that intelligent machines we build won't reenact Genesis for us,

– Tim