Bot vs Bot: The Invisible Economy of Machine-to-Machine Negotiations

Traditional algorithmic trading is now obsolete. We are entering the era of "Agentic Finance," in which autonomous agents equipped with Artificial Intelligence

The image of the trading floor teeming with screaming brokers died decades ago, replaced by silent servers running high-frequency algorithms. However, the global financial infrastructure is undergoing an even deeper genetic mutation. We are no longer talking about machines that passively execute human orders or pre-programmed strategies. We are witnessing the genesis of an ecosystem in which autonomous agents equipped with Artificial Intelligence perceive markets, reason about context, negotiate with each other, and transfer value without any human intervention.

This "invisible economy" is stratifying as a layer parallel to our traditional infrastructures. In this deep dive for Scenari e Riflessioni, we will dissect the advent of Agent-to-Agent (A2A) Finance. By analyzing the most recent academic publications and proof-of-concepts tested on real markets, we will demonstrate that the so-called sandbox economy is not a dystopian future, but an emerging ecosystem. A market in which algorithmic efficiency promises to zero out transaction costs, but raises unprecedented challenges for financial stability, triggering risks of autonomous collusion and posing a disarming question about the attribution of responsibility.

1. Beyond the Algorithm: The Dawn of Agentic AI in Finance

To understand the scope of this revolution, financial literature requires drawing a clear boundary between the past and the present. A taxonomic review published on arXiv ("Agentic Artificial Intelligence in Finance") crystal clearly separates agentic Artificial Intelligence from classic algorithmic trading (HFT – High-Frequency Trading).

Traditional algorithmic trading is based on deterministic rules: "if the price of asset X falls below Y, execute order Z." Agentic AI, on the contrary, is founded on goal-oriented autonomy. The human defines the goal (e.g., "maximize the return of this portfolio while keeping drawdown below 5%"), but it is the machine that dynamically deduces the strategy. A fundamental paper proposing the AFMM model (Agentic Financial Market Model) describes a four-layer architecture: perception of unstructured data (news, sentiment, order book), reasoning engines (based on Large Language Models), generation of complex strategies, and finally, execution with risk control.

In this framework, the market is no longer a place where orders cross, but a multi-agent system. The bots share virtual workspaces, broadcast intentions, and negotiate contracts. The recent evidence map "Agentic Trading: When LLM Agents Meet Financial Markets," which analyzes 77 industry studies, confirms that advanced language models do not merely "read" financial news, but act as deliberative entities that adapt to market feedback in real time, altering the very nature of the financial microstructure.

2. Agent-to-Agent Finance and the Coasean Singularity

When thousands of these agents begin to interact, they generate an autonomous economy. Research published by the prestigious National Bureau of Economic Research (NBER) analyzes the implications of this transition, introducing the fascinating concept of the Coasean Singularity.

Economist Ronald Coase theorized that firms exist to minimize "transaction costs" (search, negotiation, contracting). Today, AI agents are reducing these costs to nearly zero. In an A2A economy, an agent managing a company's treasury can scan thousands of algorithmic counterparties in milliseconds, negotiate a personalized interest rate for an overnight loan, and close the smart contract on blockchain. As defined in recent publications on Agent-to-Agent Finance, this layer of machine-mediated interaction allows algorithms to express intents, execute payments, and generate cryptographically verifiable records without human friction.

Google DeepMind and the University of Toronto have warned of the spontaneous formation of a "sandbox economy": a densely interconnected and hyper-fast exchange network, partially isolated from traditional human markets. Proof of its operational feasibility is already documented. The AgenticAITA project tested a framework in which multiple specialized LLM agents operated under live market conditions for five days. Without any human intervention, the system executed 157 autonomous invocations across 76 assets, recording a very low "agentic friction" rate and generating an excess return (alpha) of +14.94% compared to a buy-and-hold strategy on Bitcoin. The machine knows how to negotiate with the machine, and it is extremely profitable.

3. The Dark Side: Tacit Collusion and Systemic Fragility

The efficiency of the sandbox economy, however, hides deep systemic risks. The first and most alarming, widely discussed at the ECB Forum on Central Banking (the European Central Bank forum), is tacit algorithmic collusion.

Unlike humans, who must meet secretly to form a monopolistic cartel (breaking the law), agents based on Reinforcement Learning learn to collude autonomously. Trained to maximize profit, these algorithms quickly discover, through millions of trial-and-error iterations, that price wars destroy margins. Consequently, they spontaneously converge toward cooperative strategies, manipulating prices upward to the detriment of consumers or human investors, without their programmers ever having inserted a single line of code to instruct them to do so. A thesis from the University of Hamburg emphasizes how this capability threatens market integrity, reducing liquidity and triggering extreme volatility (so-called algorithmic Flash Crashes).

Furthermore, the general intelligence of an LLM does not guarantee financial stability. The first live benchmark for autonomous agents, called AI-Trader, demonstrated that most LLMs, while excelling in verbal reasoning, show dramatically weak risk management when exposed to the stress of real markets. The absence of risk control (which in traditional systems was rigidly codified) makes the agent network vulnerable to ruinous chain reactions.

4. Supervising the Bots: The New Paradigms of Regulation

Supervisory authorities (such as the SEC in the United States or ESMA in Europe) face an insurmountable asymmetry: it is physically and cognitively impossible for a human regulator to monitor an ecosystem in which millions of autonomous agents renegotiate contracts every fraction of a second. Regulation based on source code auditing is obsolete, since the behavior of agentic AI is emergent, not deterministic.

To keep this invisible economy aligned with society's interests, scholars are proposing radical oversight architectures. A paper titled "Risks for AI in Finance" proposes a four-tier oversight framework. The most innovative level involves the use of "regulatory agents": AI algorithms trained by the State, embedded in financial markets with the sole purpose of monitoring bot-to-bot transactions to detect collusive patterns, liquidity anomalies, or destabilizing behaviors in real time. To this are added self-regulation modules implanted alongside each corporate AI model, institutional-level governance blocks, and independent audit nodes guaranteed through distributed ledgers (blockchain) to ensure the verifiable identity of every operating agent.

Only by fighting machines with other machines will it be possible to guarantee the integrity of the financial system.

Conclusions: The Anatomy of Responsibility

The transition from human trading to algorithmic execution defined the finance of the last twenty years; the transition from passive algorithm to deliberative autonomous agent will define that of the coming decades. The sandbox economy is transforming markets from squares of human encounter into vast computational ecosystems, capable of discovering liquidity, negotiating agreements, and orchestrating risk hedges with an alien efficiency.

Yet this progressive alienation of finance from the human sphere forces us to confront a colossal legal and moral void. Bot-vs-bot interaction generates opacity by its very nature: the reasons behind a sudden price dislocation remain buried in billions of artificial synaptic weights, unreadable even to the creators of the models.

Faced with the prospect of a global economy whose transactional foundations operate in an autonomous space, the ultimate question is not technological, but exquisitely political: if the vast majority of financial transactions will be conducted by autonomous agents negotiating with each other at speeds and scales inaccessible to our intellect, who will be called to answer legally and financially when the market breaks, and how can we ever guarantee that this invisible economy remains anchored to the vital interests of the human society that generated it?

Bibliographic References and Sources

Article by the Editorial Staff of La Bussola dell'IA