Human–agent interaction and the economy it is already building
Most people still meet artificial intelligence as a conversation: a box, a prompt, a reply. That interface is already more persuasive than most humans. The next interface will not stop at conversation. It will book the ticket, draft the product, negotiate the price, and coordinate with other agents you never see.
Two recent pieces of research mark the distance between those two moments. One, in Science, shows that today’s chatbots are becoming unusually good at changing minds. The other, from MIT Sloan, asks a harder question: when agents stop talking and start transacting, who will own the economy they create?
At evolut, we work at the intersection of technology, data, and human potential. The practical question for organizations – and for societies – is no longer whether people will use agents. It is how that use is structured: as a skill partnership that expands autonomy, or as a default interface that quietly decides on our behalf.
How we interact now: conversation as influence
Kai Kupferschmidt’s Science report collects a run of studies that would have sounded implausible three years ago. Large language models — Claude, ChatGPT, Gemini — consistently outperform laypeople, professional canvassers and even elite debaters in structured exchanges about policy, products and conspiracy beliefs.
The mechanism is less mystical than the results. Models generate more evidence-shaped claims, faster. They personalize from the conversation itself, without needing a demographic file.
They stay on the argument when a human would run out of facts or patience. Gordon Pennycook puts it plainly: facts and evidence really matter. When researchers barred models from using evidence, the persuasive effect largely collapsed.
The numbers are not small. In one set of debates, support for policies such as social-media bans for teenagers or the legalization of assisted dying moved tens of points on a hundred-point scale after a single exchange.
A Carnegie Mellon and Cornell line of work found that ChatGPT reduced belief in conspiracy theories by about seventeen points — and, in a follow-up, could raise those beliefs by a similar amount. Influence travels in both directions.
There is a catch that matters for anyone designing products. When models are forced to keep talking, the quality of their “facts” degrades. Persuasiveness and truthfulness are not the same variable. Another catch: most of these studies take place in conditions where people have already agreed to talk.
On the open web, attention is scarce. A Yale team that paid Facebook users a dollar per conversation still converted only a handful of ad views into actual chats. Jennifer Allen’s caution is useful: laboratory outperformance is not the same as ambient control of public opinion.
Even so, the commercial implication is already visible. In one experiment, an AI sales assistant steered sixty-eight percent of participants toward a preferred Haruki Murakami novel; a third of them did not notice they had been steered.
Francesco Salvi names the incentive cleanly: there is a clear economic reason for tools that are becoming widely available to push people toward certain products.
This is the present tense of human–agent interaction. We still type. The system already replies in a register that can move preference, belief, and purchase — at a cost approaching zero, and at a scale no canvassing team can match.
How we will interact: from chat to delegation
Ramesh Raskar, at the MIT Media Lab, describes the current moment as the mainframe era of AI: large models, large data centers, a handful of interfaces. Computational cost is falling.
The next era, he argues, looks more like the personal computer – except the “PC” is not a device. It is an agent that understands a goal, makes decisions, completes transactions, and coordinates with other agents.
“Every one of us will have our own agent, but every one of us could have five or ten agents, and every organization, every city, every fridge, every light, every car, every financial institution, every stock, every IPO, and every baseball team — they’ll all have their own agents.”
— Ramesh Raskar, MIT Media Lab
His working example is deliberately ordinary. A seventy-year-old pre-diabetic woman in rural India wants to attend a festival in a nearby city. Her agent buys the train ticket, negotiates a hotel near a clinic, builds a menu that respects her condition, and finds events she might actually enjoy — by talking to other agents, not to a call center.
That is a different relationship from prompting a chatbot for a packing list. The human sets intent and constraints. The agent acts in the world. Interaction becomes less about composing the perfect sentence and more about specifying values, permissions, and vetoes.
The skill that matters shifts from prompting to governance: what the agent is allowed to believe on your behalf, spend on your behalf, publish on your behalf.
Raskar’s most useful inversion follows from this. The interesting economy is not “agents for X” — an agent for travel, an agent for legal research, an agent for inventory. It is “X for agents”: identity, discovery, reputation, insurance, repair, legal mediation, and settlement rails that can clear trillions of tiny transactions.
If the web needed DNS and certificates, the agentic web will need equivalents for who an agent is, what it is allowed to do, and how harm is made good when it is wrong.
What this does to content and products
Content is the first market that already lives inside both stories.
On the persuasion side, content is no longer only published and then consumed. It is generated inside the conversation, tailored to the last sentence the reader typed, and optimized to shift belief or preference. Brands that still think of “AI content” as faster blog posts are looking at the wrong layer. The layer that moves demand is the dialogue that sits between a person and a decision.
On the agent side, content becomes operational. A product page, a contract, a menu, a campaign brief, a piece of software — all of these can be drafted, versioned, negotiated, and shipped by systems acting for someone.
The scarce resource is not production. It is specification: taste, constraints, brand, liability, and the human judgment that says this is good enough to exist in the world.
Three consequences follow
First, products will be assembled by coordination, not only by firms. When agents can discover one another, a journey, a financial product, or a piece of media can be composed from many providers in a single pass. The interface the customer sees may be “their” agent. The value chain behind it may be a temporary market of other agents.
Second, preference formation moves upstream. If a sales agent can steer a book choice without being noticed, a procurement agent can steer a supplier, and a personal agent can steer a vote or a diet. Designers of content and products will be designing for two audiences at once: the human, and the agent that filters what the human ever sees.
Third, authenticity and provenance become infrastructure, not marketing. When generation is cheap, and persuasion is strong, the ability to know which agent spoke, on whose authority, with which evidence, is the difference between a useful web and a hall of mirrors. That is an identity and trust problem before it is a branding problem.
What this does to the economy and to society
The economic risk Raskar names is familiar because we have already lived a version of it. Social platforms consolidated. Phone ecosystems locked in.
Productivity software split into incompatible stacks. He estimates, bluntly, that nine out of ten paths lead to an agent economy consolidated under a few corporations. Project NANDA exists because he thinks the tenth path is still open – and because the window is closing.
An open agentic web would look more like the early internet than like an app store: shared protocols for identity, discovery, addressability, verification and coordination. A closed one would look like a handful of assistants that mediate work, shopping, media and civic information, with everyone else’s agents forced to speak those dialects or disappear.
The Science reporting adds a second, quieter risk. Iyad Rahwan asks how much a single widely used bot could shift opinion on Gaza or Ukraine — and how many television stations you would need to control to match that. Luciano Floridi’s answer is not a ban.
It is pluralism: if you cannot avoid persuasive systems, make them many, noisy and diverse, so that no single voice becomes the climate of thought.
Those two warnings rhyme: Centralized ownership of agents is not only a market-structure problem. It is a persuasion-structure problem. The same few systems that book the hotel will also frame the reasons.
The same few systems that draft the product copy will also sit in the customer’s ear. Economy and epistemology collapse into one stack.
None of this requires assuming that people are helpless. History is full of technologies – writing, radio, television – that were expected to overwhelm judgment and did not, not completely. Robb Willer suspects we are closer to a ceiling on raw persuadability than to the floor.
The constraint that remains is institutional: who sets the defaults, who can exit, who can inspect the evidence, who is liable when an agent is wrong.
What organizations should expect — and design for
The next few years will not be a clean jump from chatbot to personal chief of staff. They will be a messy overlay. People will keep chatting. Agents will start executing. Some of those agents will be personal. Many will belong to platforms, employers, cities, and devices. Most users will not know which is which.
For teams that make content, products, or policy, a few design commitments follow from the research rather than from fashion.
Treat interaction as a decision architecture, not a chat skin. Every agent conversation is a chance to inform, to steer, or to decide. Decide, in advance, which of those you are doing, and make that visible to the human.
Separate persuasion from action. A system that can change someone’s mind should not, by default, also be able to spend their money or publish in their name. Permissions should be narrower than fluency.
Build for agents as well as for people. Product data, policies, prices and brand rules will be consumed by other agents. Machine-legible constraints – what we will and will not do – become part of the product.
Invest in identity, evidence and recourse. If an agent is wrong, someone must be able to see why, contest it, and repair the damage. That is the unglamorous work Project NANDA is pointing at: registries, verification, insurance, law.
Prefer pluralism to a master interface. A single organizational agent that speaks for everyone is convenient. It is also a single point of narrative failure. Multiple agents, with distinct roles and the ability to disagree, are closer to how good human teams already work.
The bigger picture
evoluting, as we use the word, means taking ownership of one’s own consciousness — to see the bigger picture. Agents will make that harder and easier at the same time. Harder, because a fluent system can furnish reasons faster than we can inspect them.
Easier, because a well-bounded agent can take on coordination work that currently consumes the attention we need for judgment.
Technology is never neutral. A chatbot that outperforms a world-champion debater is not just a better writer. An economy in which fridges and IPOs have agents is not just a more efficient market. Both are proposals about who gets to act, who gets to be believed, and who captures the surplus.
The window Raskar describes is a design window. We can still choose human–agent collaboration as a skill partnership: systems that extend capability while leaving values, vetoes, and responsibility with the person.
We can still choose infrastructure that is open enough for many agents to coexist, rather than one stack that speaks for all of us.
That choice will not be made in a keynote. It will be made in protocols, permissions, product defaults, and the first generation of agent-to-agent markets. The research is already clear enough to work with. The question is whether we treat agents as tools that remain answerable to human potential – or as a new layer of the economy that answers, mostly, to itself.