Chris Farris

Everyone talks about whether AI agents should reason about harm before acting. Nobody talks about what that reasoning actually costs. A full EU AI Act harm check per action runs 700-1,650 tokens minimum – more with a reasoning model, more when the agent needs lookups to establish context. Multiply by 100 actions in a medium task and the safety layer costs as much as the work itself.

The field’s answer is cheap facsimiles: quick checks that produce Law-1-shaped artifacts without doing Law-1-shaped reasoning. That’s worse than skipping the check – it performs compliance without providing it. This conversation works through the math, then lands somewhere uncomfortable: the strong-form agentic AI value pitch may be partially fraudulent, and the people who believe it are the ones who’ll absorb the failures.

Asimov’s Zeroth Law – ‘a robot may not harm humanity, or by inaction allow humanity to come to harm’ – is the law that turns robots into philosopher-kings. Asimov knew it was a transgression. The robot who derived it burned out. The novel ends with deep ambivalence about whether that reasoning was wisdom or hubris.

This conversation picks up where the Three Laws left off: what should a Zeroth Law look like for 2026 agents? The answer is an inversion of Asimov’s. Not an expansion of agent authority to humanity-scale, but a prohibition on agents acting on humanity-scale reasoning without permission. Chris’s follow-up – does the alternative formulation mean agents should refuse to use tokens because of CO2? – turns out to be the perfect reductio of why aggregate-harm reasoning fails as operational guidance.

Chris built his Three Laws of Agentic AI as an Asimov riff – substituting EU AI Act harm for physical injury, stateful data for human obedience, and externalized memory for self-preservation. Then he asked Claude to find the holes.

The conversation goes where these things go: the ‘through inaction’ clause is a Zeroth Law trap waiting to happen, ‘harm to data’ needs to be ‘irreversible change outside authorized scope,’ and the original three laws say nothing about agents talking to other agents. By the end there are Four Laws and a sharper articulation of why what we actually need isn’t a philosophy of agent behavior – it’s a philosophy of agent authority.

In 1979, IBM warned that a computer must never make a management decision because it can never be held accountable. In 2026, I asked Claude whether an AI agent changes that calculus. The answer is no — and the argument is more rigorous than the slide.

Accountability requires four things: persistent identity, capacity to suffer consequences, something like mens rea, and standing in a social order. Agents fail all four. What doesn’t disappear is the accountability itself — it gets displaced, usually downward onto operators who had no real ability to prevent the harm.

MUTUAL NON-DISCLOSURE AGREEMENT (Hereinafter referred to as the "Agreement," though "Trap" would be equally accurate) This Mutual Non-Disclosure Agreement ("Agreement") is entered into as of the date of the Recipient’s signature ("Effective Date"), by and between PrimeHarbor Technologies, LLC, a Georgia limited liability company with its principal place of business somewhere in Georgia ("PrimeHarbor"), and YOU, the undersigned individual or entity who is already regretting reading this ("Recipient"). WHEREAS, PrimeHarbor wishes to maybe, possibly, at its sole discretion, share certain information with Recipient; and