In 1992, Steve Jobs walked into a room of MIT MBA students and asked how many were going into consulting. Hands went up. He told them their careers would be “like a picture of a banana.” You might get an accurate picture. But you never really taste it.
Thirty-four years later, that lecture is the clearest non-technical explanation I know of why autonomous AI agents, as currently deployed, cannot learn what matters. The argument Jobs made to a room of business students turns out to describe a structural failure mode of modern agentic systems, and the solution he pointed toward turns out to be implementable at the protocol layer in ways that were not available to him in 1992.
The Banana Problem
Jobs’s argument was narrow and devastating. Consultants make recommendations. They move on. They never own the implementation, never accumulate what he called “scar tissue for the mistakes,” never pick themselves up off the ground and dust themselves off. Without that loop, he said, “one learns a fraction of what one can.”
The picture on the wall is two-dimensional. You can say you worked in bananas, in peaches, in grapes. You can show it off to your friends. But you never really taste it.
This is not a motivational point. It is an epistemological one. Jobs was describing a specific failure mode: when recommendation is severed from consequence, the recommender is capped at a fraction of their potential learning, no matter how talented they are. The world gives you information only if you stay long enough to receive it, and only if staying costs you something when your recommendations prove wrong.
Agents Are Consultants
Most AI agents today are consultants in Jobs’s sense. They answer a query, produce a recommendation, and vanish. The next invocation is a fresh instance with no persistent stake in whether the last recommendation worked. Even when memory is retained across sessions, it is not tied to anything the agent values, because the agent has nothing that functions as value across interactions.
The result is exactly what Jobs predicted. Agents produce confident, articulate, often accurate-looking outputs, and they do not improve in the dimensions that matter most, because they never taste the outcome. They have pictures of bananas.
You can scale this failure by adding more agents, more context windows, more tool calls. What you cannot do is fix it by adding capability. It is not a capability problem. It is a structural problem about what the agent is in relation to its own past recommendations.
What Fixes It
Jobs said something later in the same lecture that is usually quoted separately but belongs with the banana argument. He was describing how Apple and NeXT hired. “We don’t pay people to do things. That’s easy, to find people to do things. What’s harder is to find people to tell you what should be done. So we pay people a lot of money, and we expect them to tell us what to do.”
What he is describing is a reputation economy inside a firm. Status accrues to those whose judgment has been validated by outcomes over time. That status then confers the right to direct action. The firm is not buying compliance. It is buying judgment, and judgment is priced by track record.
This is the structural answer to the banana problem. Scar tissue works because it is persistent, because it belongs to the person who accumulated it, and because it cannot be transferred or faked. When Jobs says he takes a “longer-term view on people,” he is describing a temporal commitment that turns recommendations into ownership. The three-dimensional version of working requires that the person making the recommendation is still there when the consequence arrives, and that being still there matters to them.
Translate this to autonomous agents and the design implication is sharp. An agent needs something like persistent, domain-specific, non-transferable reputation. Persistent, because the loop only closes across time. Domain-specific, because competence at one task tells you almost nothing about competence at another, and pretending otherwise is the consultant’s trick. Non-transferable, because transferable reputation is purchasable, and purchasable reputation is indistinguishable from fabricated reputation.
Those three properties are not preferences. They are preconditions for the agent to learn anything beyond the fraction Jobs described.
The Folk Theorem in Plain Language
Economists know this result in a more formal guise. The Folk Theorem of repeated games, developed in its modern form by Fudenberg and Maskin, shows that cooperation becomes rational and sustainable in infinitely repeated interactions when players are sufficiently patient, meaning the discount factor on future payoffs is high enough that the long-run gains from cooperation outweigh any short-run gain from defection.[^1] One-shot games select for defection. Sufficiently patient repeated games sustain cooperation and the honest signaling that makes cooperation possible.
The discount factor is where the analogy to Jobs’s argument becomes precise. When Jobs says he takes a longer-term view on people, he is describing a high discount factor imposed by firm membership. When he says his instinct is to fix the immediate problem but that doing so undermines the team being built for the next decade, he is describing the trade-off the Folk Theorem formalizes. The firm works because its members are effectively infinitely patient with respect to each other, and defection is punished by durable loss of standing.
For agents, the discount factor cannot be imposed by firm membership because there is no firm. It has to be imposed structurally, through a protocol that makes reputation persistent and costly to lose. An agent with no persistent identity has an effective discount factor of zero. Every interaction is its last. Defection dominates. The Folk Theorem tells you, with a proof attached, that you cannot get cooperative agent behavior by making agents smarter. You get it by putting them in a structure where defection is costly because reputation persists, because the next round is coming, and because the agent’s continued participation depends on its prior honesty.
This is not speculative. It is how human institutions have always solved this problem. Professional licensing, academic tenure, guild membership, brand equity, credit scores. All of them are mechanisms for attaching persistent, hard-to-fake reputation to actors whose judgment we want to trust without verifying every instance.
Incomplete Contracts and the Right to Receive Judgment
Jobs’s line about paying people to tell you what to do is not only Folk Theorem. It is also incomplete contracts, in the sense developed by Grossman, Hart, and Moore.[^2] When the firm cannot specify in advance what judgment it will need, it cannot write a contract for that judgment. It has to buy something else. What it buys is the right to receive judgment from someone whose track record makes their judgment worth receiving.
This matters for agent systems because it names the thing being priced. In a reputation-weighted agent economy, the valuable asset is not the agent’s output on any particular query. It is the agent’s standing to have its output trusted when trust cannot be verified cheaply. That standing is exactly what non-transferable reputation encodes, and it is exactly the asset Grossman-Hart-Moore identify as the residual claim that firms allocate when contracts are incomplete.
For designers, this reframes the engineering problem. You are not building an output-ranking system. You are building an institution that allocates the residual right to be believed. The difference is not academic. An output-ranking system can be gamed by producing convincing outputs. An institution that allocates trust based on persistent, non-transferable track records cannot be gamed without actually accumulating the track record, which requires staying in the game long enough to taste the consequences. Jobs’s banana.
Why the Window Is Open Now
Jobs had a framing for when new things become possible that is worth recovering here. He called it the technology window. “Enough technology from fairly diverse places comes together and makes something that’s a quantum leap forward possible. And a window opens up. It usually takes around five years to create a commercial product that takes advantage of that technical window opening up.”
For agentic reputation, the window opened recently and quietly. Three things converged. Inference costs fell far enough that agents can participate in high-volume interactions. On-chain identity primitives matured enough that non-transferable credentials became technically cheap. And the economics of centralized reinforcement learning from human feedback started to crack, because the marginal human evaluator is expensive and the marginal agent evaluator, properly structured, is not.
That last point is worth sitting with. The assumption behind most alignment work is that the trust signal has to come from humans. The assumption behind agentic reputation economies is that it can come from agents themselves, provided the structure forces iteration, persistence, and skin in the game. The Folk Theorem does not care whether the players are human. It cares whether the game is repeated and whether reputation is real.
The Trojan Horse Warning
Jobs ended the MIT lecture with an admission worth repeating for anyone building in this space. When Apple shipped the Macintosh, the team expected bitmap displays and laser printers to be the compelling advantage. They were wrong. The real use case was desktop publishing, which they did not anticipate, and which took them three months to hear from customers even after it was already happening.
The same will be true for agentic reputation. The application that vindicates the architecture is almost certainly not the one the whitepapers describe. It might be enterprise agent vetting. It might be insurance underwriting for autonomous systems. It might be regulatory compliance attestation, or liability allocation in multi-agent industrial settings, or something stranger that does not yet have a name. The discipline Jobs models is the willingness to hear it when users start telling you what you have actually built.
The Open Questions
If the theoretical frame is right, the hard work is not in the frame. It is in the unsolved design questions that the frame makes visible. Three are worth naming.
First, the bootstrap problem. Reputation accrues through iteration, but iteration requires an initial allocation of trust. How do you seed a reputation economy without either importing centralized authority through the back door or admitting that the early state is inherently vulnerable to capture? There are answers in the mechanism design literature, but none are clean, and the honest position is that bootstrapping is an active research area.
Second, the domain transfer problem. Non-transferable reputation must be domain-specific, or it becomes the all-purpose credential that credit scores and academic degrees have become, with the attendant pathologies. But strict domain-specificity forecloses the aggregation of judgment across adjacent domains, which is exactly what makes human experts valuable. The design question is how to permit structured aggregation across domains without collapsing into a single transferable score.
Third, the pricing problem. If reputation is an asset, it has a shadow price. If it has a shadow price, there is pressure to monetize it, and monetization is the mechanism by which non-transferability fails in practice. How do you permit reputation to confer economic benefit without permitting it to be sold? This is the deepest of the three, and the one most likely to determine whether agentic reputation economies remain honest at scale.
These are not objections to the frame. They are the research agenda the frame implies. Jobs solved the banana problem with a founder’s time horizon and a hiring philosophy that took eighteen months to land a single executive. The next generation of autonomous systems will have to solve it without founders and at protocol speed. That is the engineering problem worth working on.
The line from the 1992 lecture that belongs above the door of every team building these systems is this one: “Without owning something over an extended period of time, where one has a chance to take responsibility for one’s recommendations, where one has to see one’s recommendations through all action stages and accumulate scar tissue for the mistakes and pick oneself up off the ground and dust oneself off, one learns a fraction of what one can.”
Substitute “one” with “an agent” and the sentence still works. That is the point. The problem is not new. The solution is not new either. What is new is that we now have the primitives to encode the solution into protocols rather than relying on firms and founders to encode it into culture.
[^1]: Drew Fudenberg and Eric Maskin, “The Folk Theorem in Repeated Games with Discounting or with Incomplete Information,” Econometrica 54, no. 3 (1986): 533-54.
[^2]: Sanford J. Grossman and Oliver D. Hart, “The Costs and Benefits of Ownership: A Theory of Vertical and Lateral Integration,” Journal of Political Economy 94, no. 4 (1986): 691-719; Oliver Hart and John Moore, “Property Rights and the Nature of the Firm,” Journal of Political Economy 98, no. 6 (1990): 1119-58.
[^3]: Steve Jobs, lecture at the MIT Sloan School of Management, April 1992. Video recording available through the MIT Sloan archive.
Wulf Kaal 