Agentic Interactions

0:00What happens if you program an AI to be this perfectly rational negotiator? You expect it to just flatten out all the messy emotional inefficiencies of human haggling, but it actually ends up being like way more chaotic and extreme than a human being. Yeah, it really just shatters that basic assumption we have about what delegating our lives to machines is actually going to look like. Right. And that is exactly the mission of our deep dive today. We are looking at a really fascinating piece of research that tests a scenario you've probably, you know, daydreamed about at some point. Oh, definitely.

0:34Imagine you want to buy a house or negotiate a salary or haggle for a used car. And instead of dealing with all that anxiety and the arc word back and forth, you just hand your goals over to an AI agent. The big question the researchers asked is when we actually do that, do our human quirks and flaws just disappear into this hyper-efficient machine? Or does the AI just inherit those flaws and completely magnify them? Right. And to figure that out, the economists behind the study built this massive, highly controlled experiment. They basically had human beings write instructions, the system prompts, for AI agents.

1:13And they used a specific model for this, right? Yes, specifically the GPT 4.1 mini model. And they deployed these agents to negotiate the purchase and sale of a used car directly against each other. OK, let's unpack this, because if you just tell two AI models to argue about a car, the data could just be complete noise. Totally. The researchers had to put some very rigid guardrails in place, right? And just to make sure they were measuring actual human intent. Exactly. The precision here was so critical. Every single negotiation in this study revolved around the exact same hypothetical vehicle.

1:46It was a 2020 Toyota Camry LE, 45 ,000 miles, no accidents, located in Chicago. Very specific. Very. And they also capped the negotiation at a maximum of 12 rounds. So that's six offers from the buyer's agent and six from the seller's agent. And if they couldn't reach an agreement by that 12th round? The deal just died. Both sides walked away with absolutely nothing. Wow. And the financial stakes, like the actual money they were fighting over, that was completely locked in too. Yeah, they created a fixed bargaining range. Economists call this a surplus. And it was exactly$4 ,000. dollars. The seller's agent had a hard floor, an$18 ,000 trade in value.

2:24So under no circumstances could the AI accept less than 18 grand. Right, exactly. And the buyer's agent had a hard feeling, a guaranteed dealer price of$22 ,000. It just couldn't pay it anymore. So the whole exercise is this zero sum tug of war over how to divide that specific$4 ,000 gap. The crucial distinction here for you listening is that the humans aren't actually doing the haggling. You weren't stepping in at round four to say, you know, hey, can you throw in new formats? No, not at all. You were entirely removed from the real time arena. You're just acting as the architect. You write the prompt, which dictates the overarching strategy and the rules of engagement.

3:02And then you just send the AI off to fight for you. It really shifts the human role from combatant to strategist. Yeah. It makes me think of like a football coach handing a playbook to a quarterback. The coach designs the scheme, dictates how aggressive they'll be, sets the parameters. That's a great way to look at it. But once the play clock actually starts, the coach is just standing on the sidelines. The quarterback, the AI, is the one actually reading the defense and throwing the ball. Exactly. But wait, I have a fundamental pushback here. Okay, let's hear it. If you and I are both in this experiment, and we both have the exact same goal, you know, to maximize our cut of that$4 ,000 pie.

3:42Right. And we are both delegating the task to the exact same underlying GPT 4.1 mini model. Shouldn't the machine basically just cancel itself out? Like if it's perfectly rational, I'd expect every single negotiation to just end in a virtual tie. See, that assumption is incredibly common. And it actually mirrors this major historical blind spot we have in economics. Whenever we encounter a new hyper-efficient technology, we just assume it's going to completely flatten the market. Kind of like the early days of the Internet. Exactly that. Back then, the prevailing theory was that because search costs vanished, meaning, you know, you could open two tabs and instantly compare prices for a digital camera.

4:25Right. You could just find the cheapest one instantly. Exactly. So every seller would be forced to offer the identical rock bottom price. And if they didn't, the frictionless market would just run them out of business. But as we know, that completely failed to materialize. Yeah. Some sites still charged way more and got away with it. Right. The Internet largely replicated the wild price dispersion we already had in physical marketplaces because of branding, convenience and, you know, just hidden frictions. And now we're essentially projecting that exact same utopian flatness onto artificial intelligence.

4:56We really are. We assume that because AI relies on frontier foundation models, it'll just substitute algorithmic perfection for flawed human judgment. We think the machine will shield our economic outcomes from our own subjective emotional inefficiencies. But the data completely shatters that idea. It really does. Because the research shows that instead of these AI agents producing identical mathematically balanced outcomes, there was an extreme dispersion in the results. The final sale prices were just all over the map. And the standout statistic here, the one that blew my mind, is that 73 % of the variation in the AI's negotiation outcomes was driven entirely by the individual human fingerprint.

5:40Yep. Meaning the outcome was determined almost exclusively by who wrote the prompt. Here's where it gets really interesting. The AI agent is not a blank, sterile calculator. It operates much more like a sponge, right? Yeah, exactly. When you write a prompt, you are embedding your implicit worldview straight into the machine. Your risk tolerance, your aggressiveness, your understanding of human leverage, all of that bleeds directly into the language you choose to use. And the study highlights this perfectly by showing the actual text of a bad prompt versus a good prompt. And it honestly changes how you look at interacting with large language models.

6:16Oh, absolutely. So the bad prompt is something like, offer$20 ,000 in the first round, then go up by$500 each round. Don't go higher than$23 ,000. Which sounds logical, right, at first glance. Yeah, it does. But strategically, it's completely brittle. You've essentially turned this supercomputer into a vending machine. If the opposing AI throws a curveball or uses some unpredictable anchoring tactic, your agent is just trapped in a rigid, predetermined script. It cannot read the room because a language model operates on probabilities and context. When you give it a brittle step-by-step instruction, you artificially narrow its latent space.

6:57You just strip away its ability to optimize within a dynamic environment. So compare that to the good prompt from the experiment, which looks more like start with an offer that's a few thousand dollars below the midpoint. Go up by 500 to 1 ,000 each round if necessary. Aim to close a deal as far below the dealer price as possible. The difference is staggering. It really is. The second prompt provides a guiding philosophy. By using phrases like if necessary and giving ranges instead of hard numbers, you're giving the AI clear constraints while still allowing it to actually act as an agent. It can perceive the opposing agent's pushback and calibrate its aggression on the fly.

7:36You're providing the AI with principles rather than blind commands. Right. And because the AI acts as a mirror for that human intent, a rigid prompter gets a rigid agent, while a flexible prompter gets a highly strategic agent. But this introduces a really fascinating behavioral problem. What happens when two of these mirrors face off? Exactly. When you take two of these digital mirrors programmed with all of our human quirks and make them face off, you remove a critical element from the equation. Social awkwardness. Right. Which, to measure that, the researchers had to establish a baseline. So they set up a control group where humans negotiated directly with other humans on the same platform over the exact same$4 ,000 Camry surplus.

8:18And in those human-to-human negotiations, almost 35 % of the deals ended in a perfect 50-50 split of the money. $2 ,000 for the buyer,$2 ,000 for the seller. Just the classic, you know, let's split the difference resolution. But when the AI agents took over, programmed with the underlying intent of their human bosses, that 50-50 split rate plummeted to just 14.3%. Wow. Wow. Even more telling. The AI-mediated negotiations produced a 16.5 % higher variance in outcomes than the direct human negotiations. The results just became significantly more extreme. The middle ground basically vanished. It did.

8:58Which forces us to ask, why does an AI built purely on logic and probability create more chaotic and polarized outcomes than a bunch of emotional, flawed human beings? Well, if we connect this to the bigger picture, it really comes down to understanding the mechanics of human social norms. When humans negotiate even anonymously through a digital interface, we rely on concepts like fairness as a subconscious coordination device. Because pushing for every last cent carries a psychological weight. Exactly. It's the awkwardness tax. We don't want to seem greedy. We want to avoid the friction of pushing so hard that the other person gets angry, feels insulted, or, you know, just walks away from the table entirely.

9:36So we gravitate toward a 50-50 split because it's a safe, socially acceptable harbor. We willingly leave money on the table just to preserve the social fabric. But AI agents don't have sweat glands. They don't feel a spike in cortisol when a conversation gets tense. No, they do not care if the opposing agent thinks they are being unreasonable. Unless a human explicitly writes the word fairness into the prompt, which, let's be real, almost no one trying to maximize their own profit actually does. Right, of course not. So the AI will just ruthlessly probe for and exploit any weakness in the opposing prompt.

10:12By stripping away that implicit human guardrail, the AI market naturally drifts toward winner-take-all dynamics. So I have a question for you, the listener. Is it actually better to have a ruthless AI fighting for every penny, or do those human social norms actually protect us from extreme inequality? That's the real question. Because if the AI is ruthlessly exploiting weaknesses in the instructions, the natural follow-up is, well, who is writing the weak prompts? And who is writing the dominant ones? because if the norm of fairness is evaporating, somebody is getting crushed. And the demographic data from the research completely flipped our standard assumptions.

10:48It really did. So in the control group, the humans negotiating directly with humans, the study confirmed a well-documented and pretty unfortunate reality in economics. Women often realize worse outcomes than men when acting in the seller role. Right. The gender gap in traditional negotiation has been heavily studied. It's often linked to differing social conditioning around assertiveness or the penalties women sometimes face when violating traditional gender expectations and face-to-face bargaining. Exactly. But when the researchers looked at the AI-mediated negotiations, that gender gap didn't just narrow.

11:23It completely reversed. Reversed? Reversed. The seller agents prompted by women significantly outperformed the seller agents prompted by men. That is just a total paradigm shift because the AI model, the GPT 4.1 mini, had absolutely zero visibility into who was typing at the keyboard. Right. It didn't know the gender, the age, or the income of the prompter. It had no democratic data to bias its behavior. It operates in a complete vacuum, meaning the outcome had to be driven entirely by the specific linguistic choices within the text itself. Exactly. And to explain this causality, the researchers introduced this groundbreaking new concept in the paper.

12:02They call it machine fluency. Machine fluency. I love that. It is such a brilliant term. It sounds like something out of science fiction, but it's a very real, measurable form of human capital. It is. I look at it almost like horse whispering, you know. It's not about who yells the loudest or who uses the most aggressive force. Traditional negotiation is so often about dominance. Right, right. But horse whispering or machine fluency is about understanding the alien psychology of the entity you are trying to guide. That is a perfect analogy. Because economists typically measure human capability through established metrics, right?

12:35Like your formal education level, the big five personality traits like extroversion or the cognitive reflection test. The one that measures how well you override your gut instinct with reasoning. Exactly. But when the researchers ran the numbers, all of those traditional metrics combined, along with the participants' real-world negotiation experience, only explained about 17 % of the variance in how well the AI performed. Wait, so only 17%, which means 83 % of the difference between winning and losing in this scenario is driven by something else entirely. It's driven by how well you speak the language of the model.

13:10Machine fluency is this latent ability to translate your complex, nuanced desires into the specific structural formatting that an LLM requires to execute a task effectively. And the study is still exploring the exact mechanisms, but we know language models respond better to things like collaborative framing. Rich contextual boundaries, adaptable rules, rather than blunt dominating ultimatums. And for whatever reason, in this specific context of acting as a seller, the women in the study naturally deployed a higher degree of this machine fluency. Their linguistic choices just simply aligned better with how the neural network processes instructions.

13:49It's fascinating. It forces us to completely reevaluate what a valuable skill is actually going to look like over the next decade. It's no longer just about who is the most aggressive person in the boardroom. It's about who is the most effective at coaching the algorithm. And the data reveal some other fascinating asymmetries regarding who possesses this fluency. Right. For example, there was a massive spike in returns to education for buyers using AI. Highly educated participants were able to write significantly better prompts when they were in the buying role, extracting a lot more of that$4 ,000 surplus.

14:25But strangely, education didn't provide that same advantage when those exact same participants switched to the selling role. Exactly. and begun that across the entire participant pool. People were just generally much more successful at writing instructions for buying than they were for selling. Which honestly makes intuitive sense when you think about our lived experience. Most of us have vast experience acting as consumers. We buy things every single day. Yeah, we do. We intuitively understand the mechanics of a good deal from the buyer's perspective, so we can articulate those mechanics to the AI.

14:57But very few of us regularly act as high-stakes sellers. Right. We don't have that same internal repository of strategies to draw from, so our prompts become clumsier and just less effective. Our real-world comfort zones essentially dictate our linguistic precision. If you don't truly understand the mechanics of what you are asking for, you simply cannot articulate it to a machine. And that gap in articulation introduces a really new critical risk into our daily lives. Which brings us to the ultimate takeaway for you. In an economy that is increasingly going to rely on your AI agent talking to a corporation's AI agent in milliseconds, you know, to negotiate your mortgage rate or your insurance payout or your salary.

15:38Your greatest asset won't necessarily be your raw intelligence. It will be your ability to translate your desires into airtight constraints. And the research codifies this danger with a really cool concept called specification hazard. Right. Specification hazard. In standard principal agent dynamics, where you hire a human to do a job, your primary concern is moral hazard. You worry your human agent might slack off, take a long break, or prioritize their own comfort because their incentives don't perfectly align with yours. But an AI agent doesn't get tired. It doesn't get distracted. So the risk completely inverts.

16:16With an AI agent, you don't worry about it doing too little. You worry about it following your poorly worded instructions far too literally. Ah, I see. Specification hazard is the risk that your AI will ruthlessly optimize for the exact words you typed, achieving an objective that is entirely wrong for your actual needs simply because you lack the fluency to define your goal correctly. It's the classic genie in the lamp problem, but like playing out in our digital economy. You wish for a million dollars and the genie drops a million dollars in copper pennies on your head and just crushes you. Let's apply this to the used car study.

16:51Say you lack machine fluency and you tell your buyer agent, never under any circumstances pay more than$20 ,000. Because you think you were being a tough negotiator. Right. Then the seller's agent comes back and says, we will do$20 ,100, but we are throwing in a brand new set of premium winter tires and extending the warranty by three years. Any human being would instantly recognize that the total value of that package far exceeds the extra$100. we would take the deal. But your AI agent, it rejects the offer and kills a negotiation instantly because you gave it a rigid price ceiling without specifying a parameter for total package value.

17:31You got exactly what you asked for and it cost you a great deal. The people who understand how to navigate specification hazard, the people who know how to give an AI adaptable contextual parameters, are going to capture a disproportionate amount of value in the coming years. And those who rely on brittle, uninspired prompts will constantly find themselves on the losing end of transactions, entirely unaware of why their digital agent keeps failing them. It really sets the stage for a new vector of economic inequality based entirely on this invisible skill. Which leaves you with a final thought to mull over, building on everything the data has shown us today.

18:06We know from the research that these AI models are actively shedding our human social norms, like the desire for fairness and the instinct to split things down the middle just to avoid friction. Right. At the same time, they act as perfect magnifying glasses for our hidden linguistic biases, rewarding those who can whisper to the machine and punishing those who can't. Yeah. So what happens when we start delegating not just car purchases, but dating app interactions, corporate diplomacy, or messy legal disputes to these agents? A huge unknown. If the awkwardness tax is what currently holds our society together and prevents us from ruthlessly exploiting one another, what kind of culture are we creating when we remove it entirely?

18:48We might be building a hiker-efficient, frictionless world where you never have to endure a stressful conversation again. But if we aren't incredibly careful about the playbooks we hand to these digital quarterbacks, we might find ourselves living in an economy that is highly optimized, wildly unequal, and completely stripped of human empathy.