Echoes of the future

There is a clear reason why we talk about AI agents all the time: the dream of magical creatures who are working for us is a mesmerising dream. Every AI powerhouse talk about AI travel agents that organize our trip to somewhere, order us Uber and breakfast. The vision is clear: a future where smart software agents act as digital coworkers, booking flights, managing schedules, and talking to each other to get things done.

Yoav Shoham of AI21 Labs warns in his recent article, “Don’t let hype about AI agents get ahead of reality” that we may be getting ahead of ourselves. The challenges he points out: vague definitions, unreliable systems, poor coordination, and conflicting goals, are not just minor hurdles. They are fundamental problems that threaten to derail the entire project.

What's fascinating is that these "new" problems are not new at all. They are the same core issues that Shoham and his co-author, Kevin Leyton-Brown, detailed in their 2008 classic textbook, Multiagent Systems: Algorithmic, Game-Theoretic, and Logical Foundations. As someone who has managed teams of consultants, analysts and developers for over two decades, I see the same patterns. The challenges of AI agents are the timeless challenges of managing any system of intelligent, autonomous beings. The wisdom we need for our new machines is, in many ways, quite old.

The agent's true nature

Shoham's first warning is about the word "agent" itself. He cautions against “agentwashing,” where companies sell simple automation as advanced AI. This creates a gap between hype and reality that leads to disappointment. His textbook offers a clearer path, defining a multiagent system as a collection of autonomous entities with “diverging information or diverging interests, or both” (Multiagent Systems, p. xv). This simple definition focuses on the essential traits of autonomy and potential conflict, not on the underlying technology.

This distinction is vital. In a human team, a junior analyst following a script and a senior partner making strategic decisions are both called "analysts," but they operate with vastly different levels of autonomy. Recognizing this difference is key to managing them effectively. We need, as Shoham says, “clearer expectations about what these systems are supposed to do.” It all starts with understanding the true nature of an agent.

The next major challenge is reliability. Today’s AI agents, powered by large language models (LLMs), can be unpredictable. They can invent facts, as seen when a support AI created a fake company policy. Shoham argues that we must stop treating LLMs as standalone tools and start “building complete systems around them.” This means adding checks, balances, and safety measures to ensure their outputs are accurate and safe.

This is exactly how we manage human experts. A financial analyst, like an LLM, works with incomplete data and can make mistakes. A good organization surrounds that analyst with a system of peer review, data verification, and structured analysis to catch errors. The problem is the same for both humans and AI: how do you ensure reliable results from an autonomous agent whose inner reasoning is not fully transparent? The answer lies in building robust systems, a process the book formalizes through theories of belief revision (Multiagent Systems, p. 442).

The language of coordination

For AI agents to be truly useful, they must work together. Google’s A2A protocol claims to be a universal language for agents, but Shoham points out a critical flaw: it defines how agents talk, but not “what they actually mean.” An agent might not understand the context of another's message, making coordination brittle. This isn't a new problem in computing; it's a fundamental challenge of communication.

The solution is not just a better protocol, but a shared understanding, what Multiagent Systems calls social laws and conventions (Multiagent Systems, p. 44). These are rules that constrain agent behavior to ensure they can work together effectively. Without them, even the simplest collaboration can fail.

I have seen this firsthand in business. Two teams can agree to communicate daily but still fail if they don’t have a shared definition for key terms like "project complete" or "customer success." Creating this shared context is a core task of any leader. It is about establishing a common language of meaning, not just a channel for messages. For both human and AI agents, effective coordination is built on this shared semantic foundation.

The game of self-interest

The biggest challenge of all is that agents are not always cooperative. As Shoham notes, in the real world, agents will have their own goals and represent different, often competing, interests. A travel agent working for me may get a biased quote from an airline agent whose goal is to maximize profit for the airline, not to find my cheapest flight.

To expect agents with conflicting goals to cooperate seamlessly is, as Shoham puts it, “wishful thinking.” This is the central premise of noncooperative game theory, the main subject of his book. Agents are modeled as self-interested actors trying to maximize their own "utility" (Multiagent Systems, p. 67). The solution is not to wish for selflessness, but to design the system so that cooperation becomes the most rational choice.

This is the field of mechanism design, which the book calls "inverse game theory" (Multiagent Systems, p. 274). Instead of predicting an outcome, you design the rules of the game to produce the outcome you want. This is done through "contracts, payments, or game-theoretic mechanisms." A powerful example is the Vickrey-Clarke-Groves (VCG) mechanism, which uses a clever payment system to make it rational for agents to tell the truth about their preferences (Multiagent Systems, p. 292).

This is precisely what managers of people do. A company's bonus structure is a mechanism designed to align the self-interest of employees with the goals of the company. A flawed mechanism can lead to internal competition that hurts everyone. A well-designed one fosters collaboration. The challenge is the same whether the agent is a human employee or a piece of software.

A new beginning, an old path

Yoav Shoham’s recent article is a timely warning. It reminds us that building a world of AI agents is not just an engineering problem; it is a deep, conceptual challenge. The roadmap for navigating this challenge can be found in the foundational principles of multiagent systems, which his book so clearly lays out. The book systematically covers distributed problem solving for cooperative settings, the strategic realities of noncooperative game theory, and the powerful tools of mechanism design for engineering cooperation.

The path forward requires us to look back. The problems of agency, reliability, coordination, and incentives are not new. They have been studied for decades by computer scientists, economists, and philosophers. The "thoughtful design" that Shoham calls for will not be invented wholesale. It will be built on the solid foundation of this existing knowledge. The agentic future is exciting, but it will only be realized if we learn from the past. The work is just getting started, and the path ahead is a familiar one.