If I were an attorney in court, I would argue that a "mechanical or automatic agent" cannot truly be a personal agent unless it can be trusted to do things only in line with that person's wishes and consent.

If an LLM "agent" runs amok and does things without user consent and without reason or direction, how can the person be held responsible, except for saying that they never should've granted "agency" in the first place? Couldn't the LLM's corporate masters be held liable instead?

So in a case like this, if your agent exceeded its authority, and you could prove it, you might not be bound.

Keep in mind that an LLM is not an agent. Agents use LLMs, but are not LLMs themselves. If you only want your agent to be capable of doing limited actions, program or configure it that way.

A guy who's not a lawyer arguing about lawyering with an actual lawyer. Typical tech bubble hubris.

It's basically the same with longstanding customer service "agents". They are authorized to do only what they are authorized to semantically express in the company's computer system. Even if you get one to verbally agree "We will do X for $Y', if they don't put that into their computer system it's not like you can take the company to court to enforce that.

It's not that straightforward. A contract, at heart, is an agreement between two parties, both of whom must have (among other things) reasonable material reliance in each other that they were either the principals themselves or were operating under the authority of their principal.

I am sure that Air Canada did not intend to give the autonomous customer service agent the authority to make the false promises that it did. But it did so anyway by not constraining its behavior.

> It's basically the same with longstanding customer service "agents". They are authorized to do only what they are authorized to semantically express in the company's computer system. Even if you get one to verbally agree "We will do X for $Y', if they don't put that into their computer system it's not like you can take the company to court to enforce that.

I don't think that's necessarily correct. I believe the law (again, not legal advice) would bind the seller to the agent's price mistake unless 1/the customer knew it was a mistake and tried to take advantage of it anyway or 2/the price was so outlandish that no reasonable person would believe it. That said, there's often a wide gap between what the law requires and what actually happens. Nobody's going to sue over a $10 price mistake.

Fundamentally, running `sdkmanager --licenses` does not consummate a contract [0]. Rather running this command is an indication that the user has been made aware that there is a non-negotiated contract they will be entering into by using the software - it's the continued use of the software which indicates acceptance of the terms. If an LLM does this unbeknownst to a user, this just means there is one less indication that the user is aware of the license. Of course this butts up against the limits to litigation you pointed out, which is why contracts of adhesion mostly revolve around making users disclaim legal rights, and upholding copyright (which can be enforced out of band on the scale it starts to matter).

[0] if it did then anyone could trivially work around this by skipping the check with a debugger, independently creating whatever file/contents this command creates, or using software that someone else already installed.

(I edited the sentence you quoted slightly, to make it more explicit. I don't think it changes anything but if it does then I am sorry)

You don't have to explicitly hold yourself out as an agent to be treated as one. Circumstances matter. There's an "apparent authority" doctrine of agency law I'd encourage you to study.

> Rather running this command is an indication that the user has been made aware that there is a non-negotiated contract they will be entering into by using the software - it's the continued use of the software which indicates acceptance of the terms.

Yup, that's a contract of adhesion, and so-called "click-wrap" agreements can be valid contracts. See e.g. https://www.goodwinlaw.com/en/insights/publications/2022/08/...

> if it did then anyone could trivially work around this by skipping the check with a debugger, independently creating whatever file/contents this command creates, or using software that someone else already installed.

Courts tend not to take kindly to "hacking attempts" like this, and you could find yourself liable for copyright infringement, trespass to chattels, or possibly even criminal charges under CFAA if you do.

Let me put it this way: U.S. and English law are stacked squarely in favor of the protection of property rights.

Yes, because law is generally defined in terms of intent, knowledge, and other human-level qualities. The attempt to "hack around" the specific prompt is irrelevant because the specific prompt is irrelevant, just like the specific weight of paper a contract is printed on is irrelevant - any contract could define them as relevant, but it's generally not beneficial to do so.

> There's an "apparent authority" doctrine of agency law I'd encourage you to study

Sure, but this still relies upon an LLM agent being held out as some kind of bona fide legal agent capable of executing some legally binding agreements. In this case there isn't even a counterparty who is capable of making that judgement whether the command is being run by someone with the apparent intent and authority to legally bind. So you're essentially saying there is no way for a user to run a software program without extending it the authority to form legal contracts on your behalf. I'd call this a preposterous attempt to "hack around" the utter lack of intent on the part of the person running the program.

The instruction prompt is absolutely relevant: it conveys to the agent the scope of its authority and the principal's intent, and would undoubtedly be used as evidence if a dispute arose over it. It's not different in kind from instructions you would give a human being.

> this still relies upon an LLM agent being held out as some kind of bona fide legal agent capable of executing some legally binding agreements

Which it can...

> You're essentially saying there is no way to run a software program without extending it the legal authority to form legal contracts on your behalf.

I'm not saying that at all. Agency law is very mature at this stage, and the test to determine that an actor is an agent and whether it acted within the scope of its authority is pretty clear. I'm not going to lay it all out here, so please go study it independently.

I'm also not entirely sure what your angle here is: are you trying to say that an LLM-based agent cannot under any circumstances be treated as acting on its principal's behalf? Or are you just being argumentative and trying to find some angle to be "right"?

That’s a hot take indeed.

At $WORK, a multi-billion company with tens of thousands of developers, we train people to never "click to accept", explaining it like "look, you wouldn't think of sitting down and signing a contract binding the whole MegaCorp; what make you think you can 'accept' something binding the company?"

I admit we're not always successful (people still rarely click), but at least we're trying.

That sounds pretty heavy-handed to me. Their lawyers almost certainly advised the company to do that--and I might, too, if I worked for them. But whether it's actually necessary to keep the company out of trouble....well, I'm not so sure. For example, Bob the retail assistant at the local clothing store couldn't bind his employer to a new jeans supplier contract, even if he tried. This sounds like one of those things you keep in your back pocket and take it out as a defense if someone decides to litigate over it. "Look, Your Honor, we trained our employees not to do that!"

At least with a mechanical agent, you can program it not to be even capable of accepting agreements on the principal's behalf.