The Abuse of the AI Agent Concept: Why "11 AI Agents" Is Meaningless

A Single Remark That Sparks Reflection

Recently, well-known technologist Boris Mann posted a brief but thought-provoking comment on Bluesky:

"11 AI Agents" as a phrase is mass meaningless. If I said "I have 11 spreadsheets" or "I have 11 browser tabs" to do my work, it would mean about the same thing.

The remark was reshared by Simon Willison and quickly sparked widespread discussion in the tech community. In an era where the AI Agent concept is everywhere, this analogy is remarkably precise — it exposes a problem that pervades the industry yet few address head-on: we're using numbers to package a concept that hasn't even been clearly defined yet.

AI Agent: An Over-Marketed Term

Quantity Does Not Equal Capability

When a company claims its product has "11 AI Agents," what useful information does that number actually convey? The answer: virtually none.

As Boris Mann's analogy reveals, saying "I have 11 spreadsheets" tells you nothing about how productive you are or what problems you've solved. What matters isn't how many tools you have, but what responsibilities each one carries, what tasks they complete, and what value they produce.

By the same logic, "11 AI Agents" could mean 11 independent automated processes, or it could simply mean a single simple task split into 11 steps, each slapped with the "Agent" label. Without context, the number is just noise.

The Root Cause: The Vagueness of the AI Agent Definition

The root of this problem is that "AI Agent" itself lacks an industry-accepted, clear definition.

The concept of an AI Agent (intelligent agent) can be traced back to "software agent" theory in 1990s artificial intelligence research. In classical AI literature, an Agent is defined as a computational entity capable of perceiving its environment, making autonomous decisions, and taking actions to achieve goals. Stuart Russell and Peter Norvig defined an Agent in Artificial Intelligence: A Modern Approach as "anything that can be viewed as perceiving its environment through sensors and acting upon that environment through actuators." However, with the rise of large language models (LLMs), the term has been co-opted, its meaning sliding from a rigorous academic definition toward much broader commercial usage. Since 2023, with the explosion of projects like AutoGPT and BabyAGI, the use of the word "Agent" in industry has grown exponentially — but each user assigns it a slightly different meaning.

In different contexts, it can refer to:

• An LLM system capable of autonomously planning and executing multi-step tasks
• A simple API call chain
• A chatbot with tool-calling capabilities
• A rule-based automated workflow

When the definition of AI Agent itself is fuzzy, measuring by quantity becomes even more absurd. It's like the era before "app" was standardized as a concept — claiming "our platform supports 100 apps" sounds impressive, but might just mean 100 different configuration files.

The Capability Spectrum of Current Agent Implementations

To understand the deeper reason why "quantity is meaningless," we need to recognize that Agent implementations on the market today span an enormous capability range. At one end of the spectrum is the simple ReAct (Reasoning + Acting) pattern, where an LLM calls external tools during its reasoning process — essentially an enhanced function-calling chain. In the middle are systems with planning capabilities, such as those using Chain-of-Thought for task decomposition that can dynamically adjust execution paths based on intermediate results. At the other end are complex systems with long-term memory, self-reflection, error recovery, and cross-session learning capabilities.

OpenAI's function calling, Anthropic's tool use, LangChain's Agent framework, AutoGPT's autonomous loop — these implementations may differ by orders of magnitude in autonomy and reliability, yet in marketing they're all called "Agents." When a simple API call chain and a self-reflective autonomous system both carry the label "Agent," saying "I have 11 Agents" is like saying "I have 11 vehicles" — which might include both bicycles and rockets. The information content of that number approaches zero.

What the Industry Actually Needs

Shifting from a Quantity Narrative to a Capability Narrative

Rather than telling users how many Agents you have, answer these far more valuable questions:

• What tasks can these AI Agents independently complete? Can they handle a customer request end-to-end, or can they only perform text transformation at one stage?
• How do they collaborate? Is it genuine multi-Agent cooperative decision-making, or just a simple serial pipeline?
• What's the degree of autonomy? How many times does a human need to intervene? How many edge cases can they handle?

These are the real dimensions for measuring the value of an AI Agent system.

It's worth noting that Multi-Agent Systems (MAS) is an important research branch of distributed artificial intelligence, with a history dating back to the 1980s. In traditional MAS research, multiple Agents negotiate, cooperate, or compete through communication protocols to solve complex problems that no single Agent could handle alone. Classic applications include distributed sensor networks, e-commerce auction systems, and traffic control systems. Current "multi-Agent systems" in the LLM context (such as Microsoft's AutoGen, CrewAI, and similar frameworks) borrow this concept but implement it in fundamentally different ways — typically multiple LLM instances playing different roles, interacting through natural language. Whether this architecture truly inherits the advantages of classical MAS or is merely a variant of prompt engineering remains a focal point of academic debate. Therefore, when companies claim to have a "multi-Agent system," the more important question to ask is: do these Agents possess genuine information asymmetry and complementary capabilities, or are they just multiple invocations of the same model?

Beware of the "Agent Washing" Phenomenon

Just like "AI Washing" a few years ago (slapping an AI label on traditional software), the industry is now experiencing a wave of "Agent Washing." Any feature with a bit of automation gets crowned with the "Agent" name; any multi-step LLM call gets packaged as a "multi-Agent system."

Agent Washing is not the first instance of term abuse the tech industry has experienced. In the mid-2010s, "Big Data" was widely misused — any product involving data processing called itself a big data solution. Then came the "AI Washing" wave — a 2018 study found that 40% of European startups claiming to be AI companies weren't actually using any substantial machine learning technology in their products. Similarly, "blockchain" was over-generalized in 2017-2018. These phenomena follow the same pattern: a concept with genuine technical value gets over-extended by marketing, causing cognitive fatigue among investors and users, ultimately harming the true innovators in the space. Gartner's Hype Cycle precisely describes this pattern — technology falls from the "Peak of Inflated Expectations" into the "Trough of Disillusionment," often not because the technology itself has failed, but because over-marketing has depleted market trust.

This approach might help with marketing and customer acquisition in the short term, but in the long run it severely dilutes the concept's value, drowning truly breakthrough Agent technology in noise. For the healthy development of the entire AI industry, this is an insidious form of damage.

Back to Basics: The Value of AI Agents Isn't in the Numbers

The reason Boris Mann's brief comment resonated so widely is that it used the most straightforward analogy to puncture an industry bubble. What the tech community needs isn't more Agent quantity competitions, but serious discussion of these questions:

1. What truly qualifies as an AI Agent? What minimum capabilities are required — autonomous planning, tool use, memory, error recovery?
2. How should we evaluate Agent quality? What benchmarks and evaluation frameworks need to be established?
3. Where is the real value of multi-Agent architectures? In what scenarios does multi-Agent collaboration genuinely outperform a single system?

In this era of rapid AI evolution, maintaining conceptual clarity and honesty matters more than chasing numbers. Next time you see a product claiming to have "N AI Agents," try asking: So what? What exactly can they do for me?