The market has always been a coordinating machine without a single center. Millions of individual decisions, prices, expectations, fears, and desires produce an order that is never entirely designed by anyone. This is precisely what has long interested economists: a system capable of organizing immense amounts of dispersed information without absolute central command.

Artificial intelligence does not replace this logic. It changes the nature of the actors involved. Until now, even when software executed a transaction, the intent and analysis were largely defined by humans. Tomorrow, a growing number of economic agents will be able to observe, compare, negotiate, and react much more autonomously. The market will therefore become a multi-agent system in the almost literal sense.

This is where the geometry of competition changes.

A human consumer cannot constantly compare thousands of offers. An agent can. A human company cannot renegotiate its prices with every customer every second. A system can. A bank cannot manually recalculate the risks of all its counterparties every minute. Models can do this almost continuously.

The first consequence is therefore a massive reduction in research and transaction costs. This is potentially very positive. A more transparent market, in which every buyer can quickly compare prices and terms, can become more competitive. The profits derived from opacity diminish. This is exactly what I experienced with Artprice. Opacity is never neutral.

Whoever possesses the information can capture value simply because others do not know. Making information accessible transforms the balance of power.

The personal economic agent could generalize this logic. The consumer no longer looks at five offers. Their agent looks at five thousand. The merchant can no longer rely as much on inertia or the difficulty of comparison. This can put considerable pressure on prices and quality. But, as always, the second-order consequence can be very different. If sellers also have agents capable of observing competitors in real time, adjusting prices, and learning buyer behavior, the market transforms into algorithmic competition on both sides.

We can then have extreme transparency and, simultaneously, a new opacity: no one understands exactly why the displayed price changes. Pricing is personalized and dynamic, to the point of taking on an almost conversational form. It is no longer simply the price of a product. It is the price offered to this person, at this moment, under a set of conditions. This already exists in some forms. AI can make it far more granular.

This personalization raises an important economic and moral question: should two people always pay the same price for the same good? Economics answers that legitimate price discrimination has long existed: student discounts, time-of-day pricing, negotiated rates. But when the system individually assesses your maximum willingness to pay, the relationship changes. The seller then seeks to capture almost all of the consumer's surplus. The agent may know that you have an urgent need, that your income is high, or that you are not price-sensitive.

The seller can therefore adapt the offer. This is rational for them. But it can create a profound sense of injustice if the criteria are opaque. This is why price transparency will likely become a major issue. It is not necessarily about demanding a single price everywhere, which would be economically absurd, but about making certain mechanisms visible. Consumers should perhaps know that a price is personalized and on which major categories of data it is based. Origin, again, this time applied to the price.

Where does this figure come from? We used to think of price as an objective piece of data displayed on a label. In an algorithmic market, it becomes a dynamic decision. It therefore has a genealogy.

Which model?

Which variables?

What competition is observed?

Price itself enters algorithmic calculation.

This changes the way we think about competition. Traditionally, regulators observe price levels, market shares, and agreements between companies. But what happens when several pricing agents independently learn that a certain price stability is more profitable? They may converge toward behavior resembling collusion without ever having exchanged an explicit message: "do not lower prices." This is an extremely interesting problem. Competition law has been built around proving an agreement or concerted practice.

But adaptive algorithms can produce emergent coordination. No one gave the order. Yet the result resembles an agreement. We encounter exactly the problem of swarms: local rules can produce collective behavior that no one explicitly decided. The question of responsibility then emerges as formidable. Can a company be sanctioned because its algorithm learned coordinated behavior? If the company could reasonably foresee the risk and allowed it to continue, perhaps.

If the outcome was truly unpredictable, the law will need to be more nuanced. We cannot abandon the requirement of causality simply because the system is complex. But neither can we allow opacity to become a refuge. Companies using certain advanced pricing systems will likely need to be required to monitor their effects. Again, the duty of algorithmic due diligence. A company cannot simply say: I did not program collusion.

It must examine what its system actually produces. This connects with the philosophy of the preceding chapters: whoever benefits from an agent must also assume a duty of oversight. But beware the risk of over-regulation. If every price change made by a model had to be validated by a human, we would destroy precisely the gains of automation. We must therefore distinguish normal individual action from problematic systemic behavior. The regulator could analyze patterns over time. This is where it will itself use artificial intelligence.

We are entering an economy in which market agents will be monitored by regulatory agents. Recursion again. This recalls financial markets, where automated surveillance already exists. But the same logic could extend to commerce, energy, transportation. Regulators will need capabilities comparable to those of the actors they oversee. Otherwise, cognitive asymmetry takes on enormous proportions. This is a general rule of power: a regulator unable to understand the architecture of the regulated industry gradually becomes dependent on that industry.

AI amplifies this problem because technical complexity increases. Therefore, we must invest in public capacity. This does not mean the state must possess the most efficient models on the market in every area. It must possess sufficient expertise and tools for auditing. This is precisely what I advocated for the algorithmic state. Regulation must have its own infrastructure.

But the algorithmic market raises another problem: speed. A violation or instability can occur in seconds. Traditional administrative procedure takes months or years. Legal time and algorithmic time operate on entirely different temporal scales. This requires distinguishing two functions of regulation: immediate response and investigation. The first requires circuit breakers and automatic limits.

Secondly, the opposing viewpoint can take time. Financial markets have already addressed this difficulty by suspending trading. When the system moves too quickly, the mechanism is temporarily halted to allow time for human judgment. This idea of a speed bump keeps recurring. It is not accidental. Civilization often involves introducing latency into systems capable of operating faster than our capacity for judgment. The algorithmic market will need speed bumps.

But where should we place them?

Too many restrictions can reduce liquidity and efficiency. Too few can allow panic dynamics. This is classic arbitrage. The flash crash provides a particularly clear illustration. Algorithms react to algorithms, orders propagate, prices collapse or surge without the fundamentals having changed to the same extent. The phenomenon can be corrected quickly.

But it reveals a crucial point: a market can produce a state that no one individually intended. This is a property of the system. AI amplifies this possibility because the agents become more complex and more numerous. Interactions can generate feedback loops that are difficult to anticipate. One model observes a decline and sells. Others observe these sales and conclude that the risk is increasing. They sell too.

The loop intensifies like a crowd dynamic amplified at machine speed. The market has always experienced panics. The algorithm does not invent them. It can compress them. This shift in timescale is fundamental. A movement that would have taken hours can now take only seconds.

The human being no longer has time to intervene. That is why market governance must be integrated into the protocol itself. We cannot rely only on a supervisor watching a screen. Automatic rules are needed. Code becomes micro-regulation. But this automation of safety can itself produce unexpected effects. Several circuit breakers may trigger simultaneously and shift liquidity toward other markets.

Complexity again. Regulations must therefore be tested as systems are tested: simulations, adversarial scenarios, extreme situations. Financial law becomes dynamic engineering. This is interesting because it shows the institution’s broader evolution. A rule is no longer merely a text. It becomes an operational parameter.

A limit of variation.

A delay.

A threshold.

This is precisely the convergence of law and code. But an essential distinction must be maintained: the parameter is modifiable.

Who chooses it?

On what grounds of legitimacy?

A threshold of 5 percent or 10 percent can produce different economic effects. This choice seems purely technical, yet it has a political dimension. We encounter here the very same problem as the algorithmic state. Technology contains values. The algorithmic market will therefore be no more neutral than the human market. It will simply have rules that are sometimes harder to discern.

That is why transparency about major structures is important. The public does not need to know every line of code of a stock exchange.

But the principles must be understandable:

In what situations does the market stop?

Who can intervene?

How are orders prioritized?

These are almost constitutional rules of the market. The comparison is not far-fetched. A stock exchange is an institution of trust. Participants agree to trade because they believe the rules are stable and applied fairly. If the infrastructure secretly favors certain players, trust disappears. This is why access to speed itself has historically become an issue. In electronic markets, a few microseconds can represent value.

The physical proximity of servers and the quality of connections create advantages. It is already a world where space and time are commodified. AI now adds cognitive quality. A faster agent, but also one better at interpreting signals, can seize more opportunities. We are therefore entering a competition where technological capital produces an almost instantaneous advantage. This raises a question of market fairness. Should everyone be guaranteed exactly the same technology?

Of course not. Competition also relies on innovation. But we must prevent certain asymmetries from arising from privileged access to rules or non-public data. This is the classic distinction between competitive advantage and insider trading. AI does not change this principle. It simply increases the ability to exploit information. A model can combine thousands of public data points to deduce something no one had seen.

Is this a legitimate advantage? In principle, yes. That is precisely the analysis. But when the data includes information obtained illegally or through privilege, the problem remains. The law must therefore continue to examine the source of the information. Again, this concept. Where does the signal come from?

The machine can produce a brilliant inference. If it rests on prohibited data, the result may be legally tainted. The difficulty is that AI can infer information close to private data from public data. We saw the problem with personal inferences. It also exists in markets. A model may be able to estimate that a company will announce a problem from satellite images of its factories, hiring movements, or maritime traffic. This is analysis of open data.

It can provide a considerable advantage. The market will therefore become a battle of automated economic intelligence. This can increase price efficiency because more information is incorporated quickly. But it can also strengthen the players who can afford the best data and the best computing power. Public information is not equal if some can process it much better. This is an old truth. A library open to all does not abolish the difference between those who know how to search it and those who do not.

Artprice was born precisely from this understanding. Value is not just access to documents. It is standardization, indexing, the ability to uncover relationships. AI generalizes this to the entire market. Competition will therefore shift towards data infrastructure. Players capable of building high-quality, proprietary corpora will have an advantage. This could create barriers to entry.

A startup can have an excellent model but not twenty years of data. A long-established company can have the data but a slower technical architecture. Competition now combines memory and compute. This is an interesting geometry. The newcomer sometimes has the technology. The incumbent has historical depth. Whoever knows how to combine the two has an advantage.

That is why data acquisitions become strategic. We experienced this with Artprice. Buying archives and collections was not accumulating paper. It was acquiring historical time. Some series cannot be recreated quickly. Time becomes an economic asset. AI increases its value because it can exploit what a human being could not traverse in full.

This is a triumph of the long term in a market obsessed with speed. Here's a paradox that particularly interests me: the fastest systems can leverage the oldest archives. Speed and depth are not opposed; they multiply. This is precisely my way of thinking. The most radical modernity can be built upon the oldest memory. We rediscover alchemy, the archive, the database.

The algorithmic market is therefore not just a market of milliseconds. It is a market where centuries of data can be mobilized in milliseconds.

But this capacity also increases the risk of historical overfitting. A model can find beautiful regularities that no longer exist. The market changes. Monetary systems change. Behaviors change.

A relationship observed for twenty years may disappear once enough actors use it. That is reflexivity. As soon as a strategy becomes known, its advantage diminishes. The market reacts to the model of the market. This is what profoundly distinguishes economics from many natural phenomena. The model does not remain external: it acts on the market.

AI amplifies this reflexivity. If a popular model detects that a particular signal announces a rise, millions of agents may buy as soon as the signal appears, driving the price up even before the anticipated event. The prediction comes true because it is believed. Or, conversely, it cancels itself because everyone has already acted. The system enters a recursive loop. This is extremely difficult to model. Models learn from a world that transforms in response to their own recommendations.

It is almost a cybernetic loop. That is why the algorithmic market will probably never be entirely predictable, even with very powerful AI. Computing power does not eliminate reflexivity; it can accelerate it. This is a fundamental limitation to the idea of a perfectly optimized market. The market is an adaptive system populated by actors who learn from each other. Every effective strategy attracts imitators and alters the environment.

Mutation again. Only the mutants survive. But here, mutation is permanent. An agent must recalibrate their strategies as others change. It resembles military co-evolution. The line between economic strategy and adversary strategy becomes thinner. Obviously, the market is not war.

Actors can cooperate and create shared value. But they also seek relative advantage. Agents will therefore learn to anticipate the actions of others. This can lead to impressive levels of sophistication.

The agent no longer asks only:

What is the best price?

It asks:

How will the other agent react if I offer this price?

We are entering into recursive strategic interactions. This is exactly game theory augmented by learning. Two systems can test thousands of strategies. This can improve negotiation. But it can also produce tactics that are difficult for humans to understand. An agent discovers that a particular sequence of proposals maximizes the outcome. Why?

Perhaps no one can explain it simply. Negotiation becomes opaque even as it becomes more effective. This opacity raises a question of consent. Should a company accept contracts negotiated according to strategies it does not understand? If the boundaries are respected, perhaps. But for significant decisions, the human advisor must at least understand the main terms. Again, autonomy is proportional to irreversibility.

A micro-transaction can be fully automated. A company acquisition should not be completed simply because two parties have agreed. The hierarchy of actions remains essential. The algorithmic market, therefore, does not mean the disappearance of humans. It produces a temporal stratification. Machines can handle the very fast, the repetitive, the micro. Humans remain more focused on the slow, the strategic, the irreversible.

But this boundary itself can shift. Some decisions we consider strategic today could be largely automated as systems become more reliable. Therefore, we must maintain principles, not a fixed list. Again, the ever-changing nature of permanence. Governance must regularly review its thresholds. But there is a significant political risk: as markets become too fast-moving to be understood directly, they may lose their social legitimacy.

The citizen sees a price move, a company disappear, a fund makes billions, without understanding the mechanisms. The economy may be perceived as a foreign machine. This is already sometimes the case. AI may intensify that feeling. A society cannot rest solely on an efficiency no one understands. Legitimacy requires a certain intelligibility. This does not mean every citizen must understand high-frequency models.

But citizens must be able to understand the rules, know that they are monitored, and know that abuses are sanctioned. That is exactly the difference between complexity and opacity. A nuclear power plant is complex. Society accepts its existence if it trusts the institutions that control it. The algorithmic market will need the same culture of safety. Perhaps finance will have to relinquish part of its imaginary of total freedom and think of itself more as critical infrastructure. It already is.

Payments, credit, and markets finance the real economy. A systemic failure can affect everyone. AI therefore reinforces the idea that a market is an infrastructure, not just a place for speculation. This justifies safeguards. But regulation must remain aware that risk-taking is also necessary. A risk-free market does not exist. If all losses are prevented, the price mechanism is destroyed.

It is therefore essential to distinguish between normal economic risk and systemic risk created by an architecture. This is a crucial distinction. An investor can lose money because they made a bad decision. That is the market. But millions of investors should not suffer a failure because a poorly designed protocol triggered an uncontrolled loop. The regulator's role is primarily to secure the infrastructure, not to guarantee every outcome. This distinction must remain clear.

Otherwise, moral hazard increases. This relates to the responsibility of capital. Those who reap the rewards must accept a share of the risk. AI must not allow for the privatization of performance and the automatic socialization of error. But highly interconnected systems create a problem: failing one player can trigger a contagion. This is the classic "too big to fail" dilemma.

AI can create new "too connected to fail" systems. A company may not be gigantic in terms of revenue, yet it can become a central cognitive hub for many others. A model provider. A data service provider. Its failure can propagate a shockwave. This is why economic graph analysis must be integrated into systemic regulation. Size alone is no longer enough.

We need to look at centrality. It is exactly the same logic as in terrorist networks or digital infrastructures. Which node connects how many actors? Systemic regulation will have to map these dependencies. This could lead to specific resilience obligations for certain providers. Again, critical infrastructure. If a model is used by thousands of financial companies, its status changes automatically.

It is becoming almost systemic. This does not mean that the state must control it directly. But it can require continuity plans, testing, and redundancy mechanisms. This is already the logic applied to other infrastructures. AI is being integrated into this.

And this immediately raises another question:

What happens if these critical nodes belong to companies of another power?

We return to cognitive sovereignty. A national market may be legally sovereign and technically dependent. French companies may use agents hosted in the United States or elsewhere. In a geopolitical crisis, what guarantees exist? The algorithmic market therefore also becomes a geopolitical market. Sanctions may cut access to technologies. States can use control of models as economic leverage.

This is already visible with semiconductors.

AI adds a new layer. A power that possesses the cognitive infrastructure can influence the economies of others without directly controlling their companies. This is why Europe, China, and the United States are each seeking to develop their own capabilities. This is not simply protectionism. It is also the recognition of a critical interdependence.

But too much closed sovereignty can fragment the global market.

We face the same dilemma: securing without isolating. If each bloc develops incompatible agents and different standards, transaction costs increase. Therefore, interoperability protocols must be maintained. The global market depends on common standards. My French agent must be able to understand the Indian agent's contract.

This requires formats, identities, and rules. The battle for standards will be considerable. Whoever's standard becomes dominant can structure a portion of commerce. This is exactly what happened with payment cards, operating systems, and internet protocols. Agents are becoming the next interface. Whoever controls the interface controls a form of power.

But there is a key difference: the agent can choose the provider itself. If personal agents truly optimize in the user's best interest, they can reduce the influence of dominant brands. They can automatically search for the best deal from a small, unknown player. This could foster competition. But large platforms can also control the agent itself and thus steer the search. This is the new bottleneck.

Power shifts from the results page to the assistant. A search engine showed several links. An agent can choose directly. This concentrates the intermediary function enormously. That is a major antitrust question. If a dominant agent makes purchases for hundreds of millions of people, its selection rule may determine the success or failure of thousands of companies. Even without malicious intent, its architecture becomes a private economic policy.

This warrants considerable vigilance. Recommendation criteria may need to be more transparent when an agent reaches a quasi-infrastructure level. But care must be taken not to require the disclosure of all algorithmic secrets, which would facilitate circumvention. It is all about balance.

Principles can be imposed: possibility of choosing other agents; prohibition of secretly favouring one's own services; independent audit.

This is similar to the historical debates about platforms. AI does not create the problem of dominant intermediation. It amplifies it because the interface becomes decision-making. This is a huge difference in degree. A platform that ranks products exerts influence. An agent who buys without revealing the full ranking decides. This is why the fiduciary mandate of personal agents becomes central.

If an agent is legally acting in my best interest, they should not undisclose the benefits to their supplier's ecosystem. This principle could restructure the market. It could create independent agents paid directly by the user, like advisors. The business model is crucial. A free service funded by sellers does not have the same loyalty as a service paid for by the buyer. We have already encountered this issue in finance with advisors. AI is exacerbating the conflict of interest.

Provenance of the recommendation again. The algorithmic market can therefore make the consumer extraordinarily powerful or extraordinarily manipulable, depending on the architecture. This is an essential bifurcation. If agents are loyal to the user and portable, competition can intensify. If they are controlled by a few integrated platforms, concentration may increase. Technology permits both futures. This confirms once again that the institution decides.

The same technological progress can produce an open or closed architecture. There is no technological destiny. There are choices regarding standards, laws, and economic models. This political responsibility is important. We must not say in ten years: it was inevitable. No.

Choices will be made. The role of law is to guide incentives without attempting to predict every outcome. This is a delicate matter. Over-intervention can paralyze a sector. Inaction can allow entrenched positions to become difficult to reverse. Timing becomes crucial.

Regulation after a proprietary standard dominates almost the entire market is far more challenging.

This is precisely the lock-in effect. Therefore, it is sometimes necessary to act early on interoperability rather than on the products themselves. This is a lesson learned from the internet: keep the doors open. If users can switch providers and take their data with them, competition remains possible even if one player temporarily dominates. Portability thus becomes a tool of competition policy. It is a powerful idea.

The algorithmic marketplace must be designed not only with rules governing pricing, but also with rules governing the mobility of identities and data. Competition depends on the technical architecture. This further illustrates the convergence of law and engineering. An "export" button can become a tool of economic policy. The great power struggles are sometimes hidden within file formats. I have seen it all my life. Whoever controls a closed format enjoys a monopoly.

Whoever mandates interoperability redistributes power. This is precisely why standards are so political. The economics of agents makes this reality even more apparent. But we should not fantasize about a perfectly fluid market where agents switch suppliers every second. Trust, quality, and reputation will continue to foster loyalty. An economic relationship is not solely about price. An agent might learn that a particular supplier is slightly more expensive but far more reliable.

It can incorporate qualitative criteria. This can improve the market by making reputation more measurable. But again, beware of a single score. Reputation is multidimensional: quality, delivery time, service. An aggregate score can mask these aspects. The agent should be able to adapt according to the user's preferences. For one user, it might be price; for another, ecological origin; for a third, delivery time. The market then becomes extremely personalized. This can create a fragmentation of prices and offers. The very notion of a single market becomes more abstract.

Two people no longer necessarily see the same commercial universe. This is already the case with recommendations. Agents will amplify this. This raises a sociological question: if everyone has their own personalized market, the collective perception of prices can become fragmented. Comparisons become more difficult. Public statistics may need to evolve. The average price index may become less representative if prices are hyper-personalized. We will need new measurement methods.

This is an important macroeconomic issue. Statistical institutions will need to collect more detailed data while protecting privacy. AI can help build dynamic indices, but it must not mask inequalities. An average price can hide the fact that certain groups systematically pay more. Again, it is about distribution. Regulators may need to examine disparities based on different profiles.

But this requires knowing what data is being used. Transparency is key. The algorithmic market makes measurement itself more complex. It is a paradox: we have more data, yet economic categories are becoming more fluid. A price is no longer fixed. A product can be personalized. How do we calculate inflation when the supply itself changes for each user?

Economists will need to develop new methodologies. This is an interesting shift because it shows that AI is not only transforming the market, but also the tools we use to analyze it. Economic statistics must evolve, and this could have enormous political consequences. Central banks base their decisions on indicators. If these indicators become less relevant, monetary policy may become less precise.

AI can obviously help us use more real-time data. We may be entering a higher-frequency macroeconomics. Central banks can track payments, online prices, and logistics flows almost instantly. This improves diagnosis. But again, more data does not guarantee better decisions. The risk is reacting too quickly to noisy signals. Macroeconomics needs to distinguish between trend and volatility.

The slowdown continues. A central bank that reacts to every fluctuation in real time can amplify instability. AI should therefore be used to understand more quickly without necessarily making decisions more quickly. We find ourselves back at the exact distinction made in the chapter on cognitive warfare: speed of calculation is not speed of action. This is a constant. The market can produce data in milliseconds. Monetary policy may need weeks of hindsight. Intelligence lies in respecting the right timeframe for each decision.

The ability to operate across multiple temporalities is an essential institutional competence. Companies operate by the second Investors by the minute or by the year. States over several years. AI tends to pull all timescales toward the instant. We must resist that attraction. Some things should not be accelerated. An industrial policy takes ten years. A financial market reacts in ten milliseconds. Confusing the two is dangerous. One function of the State is precisely to represent the long term that the market may underweight.

AI can help the market better anticipate the long term. But incentives often remain short-term. A manager might know that a climate risk exists twenty years from now and still be evaluated on their quarterly performance. The model does not correct the incentive. This is yet another limitation. Technology does not replace institutions. We need to align our time horizons.

This is a question of governance. The algorithmic market may shorten the horizon still further if agents optimize continuously. Every second becomes an opportunity. This may increase turnover and reduce capital’s patience. But the reverse is also possible. An agent can be programmed to pursue a strategy over thirty years without being influenced by fear or daily emotion. An algorithmic pension fund might have greater long-term discipline than a human being subject to fluctuations.

The duality is evident once again. AI can shorten or lengthen time depending on the objective. It is not the machine that dictates the timeframe; it is the reward function. If we demand daily performance, it optimizes daily life. If we demand intergenerational preservation, it can incorporate a much longer time horizon. This aligns with long-term state strategies.

AI could become a tool for long-term policy precisely because it can simulate several decades. But simulation remains dependent on assumptions. The longer the time horizon, the greater the uncertainty. Therefore, scenarios must be presented, not a single forecast. Again, the pluralism of futures. The algorithmic market could benefit enormously from this scenario-oriented approach. An investor would not only ask for the probable return, but also for disruptive scenarios.

This might improve resilience. But capital loves single figures. A score simplifies. That is dangerous. We need to learn how to show risk distributions and tails. AI is perfectly capable of doing this. The problem is often the interface.

We return to the presentation of uncertainty. A value of 7.2 percent seems accurate. Perhaps it covers a huge range. Financial design needs to become more honest. This is a governance issue. The machine can calculate a billion scenarios. If the screen displays only one figure, some of the intelligence is lost.

This is a crucial point. AI can either increase the complexity of our understanding or mask it behind a score. The choice of interface is political. This applies to the market as well as the state. It is almost a constant throughout this section. We are building highly complex systems, and we must decide how to make this complexity accessible to humans. Too much information is paralyzing.

Too little creates an illusion. The right level of synthesis is an art. Perhaps personal agents will become precisely that translation layer. The market produces immense complexity. My agent interprets it according to my objectives. This can make the system more accessible. But again:

I have to trust the agent.

We return to mandate and loyalty. Everything closes the loop. This is exactly the structure of the network: each chapter becomes a node connected to the others. The economy of agents is not a new, separate subject. It concentrates identity, evidence, sovereignty, work, responsibility.

This confirms for me that AI acts as a new layer of coordination, not because it constitutes an isolated sector, but because it permeates all other structures. The market itself is coordination. Their intersection is therefore particularly profound. We can even imagine that some markets will disappear in their current form because agents will trade directly.

Why display a fixed price if each transaction can be negotiated instantly?

In some sectors, the displayed price may become an indicative price. My agent sends a proposal.

The seller responds. A kind of automated global marketplace. It is almost a return to an ancient form of commerce with hypermodern tools. History is full of these loops. Fixed pricing in retail is historically relatively recent. AI could reintroduce large-scale negotiation. That would be particularly interesting.

But this can make the market less transparent for humans. If no one knows what others are paying, the asymmetry increases. Agents will therefore need to share certain statistics. We can imagine services that show: comparable buyers paid between X and Y. This restores power. Information again as a counterweight.

Artprice has transformed a portion of the Art Market by providing price benchmarks where opacity previously reigned. The agent-based economy could replicate this strategy in other sectors. If prices become personalized, databases of actual transactions become extremely valuable. Those who know what others have actually paid can negotiate more effectively.

We will therefore probably see systems emerge that are comparable, in their function, to what Artprice built for the Art Market: memories of actual prices and transaction conditions that feed agents. This data acquires strategic value. But it raises questions of confidentiality. Should a private transaction be shared? It will sometimes be necessary to aggregate, anonymize, and delay publication. Again, the trade-off between market transparency and commercial confidentiality. Financial markets already know this problem.

Some transactions must be published. Others are not immediately disclosed in order to avoid revealing a strategy. Algorithmic economics will generalize these issues. Total transparency is not always optimal. It can even facilitate collusion by allowing all sellers to immediately verify that no one is undercutting prices. This is a significant paradox. Transparency is often thought to increase competition.

In some architectures, too much transparency between vendors can facilitate coordination. Again, these are secondary consequences. There is no absolute rule: more data equals a lower price. The architecture must be analyzed.

Who sees what?

And when?

This temporal dimension of transparency is essential. Data can be published with a delay. This protects certain strategies while still allowing for audits. Financial markets already use this logic. AI will likely make these mechanisms more sophisticated. We will have deferred, aggregated, and conditional transparency. It is precisely this granularity that produces good institutions.

Slogans always fail. The algorithmic market is too complex to simply "open everything up" or "close everything down." Information flows must be built. Again, the architecture. And this raises the question of economic democracy. If agents become the primary intermediaries of consumption and investment, their governance will have a profound influence on society.

Who defines the ethical, environmental, geographical or provenance criteria they use?

A citizen can choose: to favour French companies; to reduce carbon impact; to maximize the lowest price.

The agent becomes an economic translation of personal values. This can give the consumer considerable power. Millions of agents may shift demand according to ethical preferences. But this can also fragment the economy into ideological bubbles. Some agents avoid every company associated with a particular political position. Others do the opposite. The market then becomes an instrument of polarization.

We are already seeing boycotts. Automation can make them permanent and invisible. A preference set once can exclude certain suppliers for years. This can have social consequences. there is also the risk of a fixed profile. Revision must be allowed.

Freedom is not only choosing once. It is being able to choose again.

An agent should therefore periodically check whether certain criteria are still desired. You excluded a certain type of supplier three years ago. Do you want to maintain this criterion? It may seem like a minor interface detail, but it is a matter of freedom. Permanent delegation can transform a past choice into a predetermined fate. The agent must therefore accept the possibility of change.

Again and again. This confirms for me that "Only the mutants survive" is not just a biological or entrepreneurial formula. It is also a principle of AI system governance. A system must allow its objectives and preferences to evolve. Rigidity is a form of vulnerability. The market itself survives because it mutates.

Businesses are emerging.

They disappear.

Prices change.

AI can accelerate this transformation. It can also lock in certain behaviors if standards become centralized. This is the paradox. The market remains vibrant as long as there is diversity and open entry. Competition is not just about the existing companies today; it is about the potential for new ones to emerge tomorrow. Interoperability is therefore a prerequisite for economic transformation.

If dominant agents can speak only to the major platforms, new entrants become invisible. Open protocols then constitute an innovation policy. This idea is essential.

Technical architecture can determine the contestability of the market.

It is almost an economic constitution. Here again, the standard is never neutral. It defines who can enter. The regulator should therefore examine the discovery protocols.

Can a small supplier be found by the agents?

Do you have to pay to have your indexation adjusted?

Can agents directly query open catalogues?

This is similar to the debates surrounding search engines, but with a more direct impact because the agent can make a purchase. The power of intermediation becomes transactional. This is likely one of the major future challenges for competition. Governments can impose fair access obligations on certain interfaces. But it is crucial to avoid undermining security. An agent must be able to verify that a provider is legitimate and trustworthy. Business identity standards will become important.

Origin and reputation are key. It all comes back to that. The algorithmic marketplace could therefore be much more technically open and much more cognitively filtered. Millions of suppliers are accessible. But the agent only displays a few based on specific criteria. The power of the filter increases. We can see, then, that the market of the future might be less about simply appearing on a shelf and more about being selected by models.

Search engine optimization (SEO) is changing. Companies will strive to make their offerings machine-readable, their data structured, and their evidence readily available. This can be positive: standardized descriptions, provenance, certifications, and proof of quality. But it can also create new opportunities for manipulation. Suppliers will learn to present their products in a way that maximizes their agents' scores. This is precisely the gamble of metrics.

The agent will therefore have to detect attempts at manipulation. Adversarial competition again. The market becomes a battleground of optimization versus optimization. This dynamic seems difficult to avoid. It can lead to a race for certifications and proof. The economy of trust again. Reliable suppliers will have an incentive to produce verifiable data.

An unaudited label will lose value. This can improve the informational quality of the market. It is an interesting prospect. AI can push companies to better structure their provenance data because agents demand it. A product without verifiable information may be less likely to be selected. This could promote environmental or health traceability.

But beware of small businesses that lack the resources to produce sophisticated data. Standards must remain accessible. Otherwise, machine-readable compliance becomes a new barrier to entry. Yet another trade-off. A standard that increases trust can favor large players who can afford certification. Therefore, proportionate mechanisms must be designed. This is exactly the same problem as regulating AI itself.

Fixed compliance costs encourage concentration. The law needs to recognize this. We could have certification levels based on risk. A small business owner does not need the same infrastructure as a bank. Again, classification. This repetition may seem obsessive. Yet, it corresponds to the necessary methodology in a complex world.

Granularity is a defense against adverse effects.

And it leads us to an even broader question.

If markets become algorithmic, what happens to the notion of price as social information?

Hayek had emphasized price as a signal that condenses dispersed information. But if prices are hyper-personalized and negotiated by agents, does a common price still exist that plays this role? Perhaps the system functions more with price distributions than with a single price. The signal becomes richer.

But less visible. Agents can understand it. Humans, less directly. We are entering an economy where machines could perceive certain market structures better than people. This creates a collective cognitive dependence. Citizens no longer directly understand prices. They ask their agents:

Is it expensive?

The model responds by comparing thousands of transactions. This is useful. But it also means that our relationship with the market becomes mediated. We already saw this with comparison sites. The agent could become an almost universal comparator. Whoever controls the comparison shapes part of our perception of value.

This is extremely important. It ties into the Art Market, where databases transform the perception of a price. Seeing the history changes judgment. Artprice has given market participants a mirror of the Art Market. AI is generalizing this mirror. But a mirror can be distorting. So we need to know what data is being entered. And again, the provenance.

The loop is complete.

Value. Proof. Identity. Market.

Ultimately, everything hinges on the quality of information architectures. This may be the provisional conclusion of this chapter. The algorithmic market does not eliminate the market. It reveals even more clearly that a market is an architecture of trust, information, and rules. AI accelerates flows and increases capacity. But it does not replace any of these foundations. On the contrary, it makes their design more critical. The price can be calculated by machine. The legitimacy of the rules remains human. Negotiation can be automated. Responsibility remains.

Competition can become more dynamic. It can also become closed off around proprietary standards.

It all depends on the architecture.

Finally, this automation of the market opens a broader question: when the same tools serve both market coordination and public action, the

classical boundary between market and planning itself comes under scrutiny .

What happens to the market when coordination becomes almost instantaneous?

What happens to planning when AI can simulate and adjust continuously?

And most importantly:

Who defines the purpose of an economy capable of simultaneously optimizing a huge number of variables?

The question is then less about the market than about the purpose: who defines what an algorithmic economy seeks to optimize?

This is where the boundary between market coordination and planning needs to be re-examined.