2026-06-20 · Ankina Lab

Why Did AI Start Having Job Titles?

AutoGen enabled conversations between AI agents. But conversation alone did not create an organization. As long as every agent was fundamentally the same model, the echo chamber problem remained. With the arrival of CAMEL, AI gained something new: roles.

Introduction

In 2023, a new shift began to emerge in the AI industry.

The release of ChatGPT at the end of 2022 had sent shockwaves around the world. For the first time, ordinary people were spending extended periods of time talking with AI. They asked questions. They learned. They debated. They sought advice. They wrote articles. They wrote code. They translated languages.

It was a fundamentally different experience from the AI systems that had come before.

But researchers were already turning their attention toward the next stage.

Not a single AI.

Multiple AIs.

What would happen if AIs could talk to one another? What would happen if they could collaborate? What if they could review each other's work? And what if they could function as a team?

As we saw in the previous article, AutoGen was one of the first major studies to demonstrate that possibility.

Yet AutoGen's success also exposed a new problem.

AI agents could communicate. But were they truly forming an organization? Were they genuinely collaborating?

And one critical question remained.

If every agent is ultimately the same AI, does building a team actually mean anything?

This article explores the emergence of CAMEL and Role-Playing Agents, tracing why AI systems began to adopt roles and job titles.

Ironically, the answer did not begin with cutting-edge AI technology. It began with a mechanism that human societies have been refining for centuries: roles.

The World Opened by AutoGen

Before AutoGen, most LLM interactions followed a simple pattern.

A user asked a question. The AI responded. The user asked a follow-up. The AI responded again. The conversation could continue, but the underlying structure remained unchanged: one human and one AI.

AutoGen challenged that assumption.

Multiple agents were created. Each agent could send messages. One agent could propose an idea. Another could evaluate it. A third could revise it. Humans could intervene when necessary, or step aside and allow the agents to continue talking among themselves.

This idea excited many researchers.

Most achievements in human civilization emerge through cooperation.

Companies. Universities. Research laboratories. Hospitals. Governments. Military organizations.

All of them depend on groups of people working together.

If human intelligence benefits from collaboration, perhaps AI could follow the same path. Perhaps multiple AI agents could solve problems that a single agent could not.

The idea seemed natural. And in many cases, it worked.

Planning agents. Execution agents. Review agents.

Dividing work among specialized participants often produced better outcomes than relying on a single agent alone.

Yet researchers gradually began to feel that something was missing.

The conversations worked. But were they truly organizations?

Conversation Is Not Organization

Consider human society.

Place ten people in a meeting room. Let everyone speak. Let everyone debate. Let everyone share opinions.

Does that automatically create a company?

Of course not. Nor does it create a research lab. Nor a project team.

Organizations require something more.

Roles.

Who makes decisions? Who conducts research? Who implements solutions? Who verifies results? Who carries responsibility?

Only when these functions exist can an organization operate effectively.

Conversation and organization are not the same thing.

AutoGen enabled conversation. It did not yet create organizations.

The Problem of Identical Agents

There was another problem as well.

All the agents were fundamentally the same model.

Imagine creating five GPT-4 agents. Give them different names. Give them different histories. Assign different tasks. Yet beneath the surface they remain identical.

The same training data. The same architecture. The same world knowledge. The same reasoning tendencies.

In human terms, it would be like watching five people hold a meeting, only to discover that all five are copies of the same individual.

Their responses would not be perfectly identical. Context and prompts would create variation. But their underlying tendencies would remain similar.

They would generate similar ideas. Make similar mistakes. Reach similar conclusions.

True diversity would be difficult to achieve.

The Echo Chamber Problem

This phenomenon is well known in sociology. It is called an echo chamber.

When people with similar beliefs gather together, their views become reinforced. Opposing perspectives become rare. Blind spots remain unchallenged. As a result, collective judgment can become distorted.

Something similar appeared in multi-agent systems.

One agent proposes an idea. Another agrees. A third agrees. A fourth agrees.

At first glance, this appears to be consensus. But perhaps it is merely the same model reasoning in the same direction multiple times.

That is not genuine debate. It is amplification.

Researchers began to realize that simply placing multiple agents together was not enough. What they truly needed was diversity of perspective.

Why Did Human Society Invent Job Titles?

At this point, researchers turned their attention toward human organizations.

Why do we create roles? Why do organizations assign titles?

Expertise is part of the answer. But it is not the whole answer.

Roles change how people think.

Engineers focus on implementation. Salespeople focus on customers. Lawyers focus on risk. Accountants focus on numbers. Executives focus on markets.

Everyone is looking at the same problem. Yet each person sees something different.

That difference creates discussion. It reveals blind spots. It enables conclusions that no single individual could reach alone.

A role is not merely a label. A role is a lens.

Role Theory

This idea is formalized in sociology through Role Theory.

Role Theory argues that human behavior is shaped not only by personality, but also by social roles.

Teachers teach. Judges evaluate evidence. Doctors diagnose. Military officers prioritize chains of command.

The same individual behaves differently when occupying different roles. Perspectives change. Evaluation criteria change. Decision-making changes.

Researchers began asking an intriguing question.

If this phenomenon occurs in humans, could it occur in AI systems as well?

Could the same model produce different perspectives simply by assigning different roles?

That question ultimately led AI systems toward job titles.

Can Prompts Create a Personality?

Role Theory revealed something fascinating.

Humans change their behavior depending on their roles. They behave as teachers. They behave as doctors. They behave as judges. They behave as executives.

Even when the person is the same, what they pay attention to changes depending on their position.

So what about AI? What happens if we assign different roles to the same model?

Today, this question may sound obvious. But around 2023, it was a highly important question for researchers.

Because it might provide a way to introduce diversity into multi-agent systems.

All agents are based on the same model. They share the same knowledge. They share the same reasoning ability. Under those conditions, how can different perspectives be created?

Researchers gradually began to focus on the prompt itself.

Persona Prompting

One early idea was Persona Prompting.

A persona means a personality or character profile. The idea is to give the AI a particular persona.

For example, instruct the model: "You are a historian."

The answer then begins to emphasize historical context. It explains causal relationships between events. It compares the present with the past. It discusses long-term change.

What happens if the same question is asked with a different instruction?

"You are an investor."

Now the response changes. The model looks at market impact. It looks at risk. It looks at profit. It looks at the competitive environment.

The model is the same. But the output changes.

Researchers saw an important discovery here. Simply assigning a persona could change behavior.

Character Prompting

Eventually, researchers realized that not only occupations but also personality traits could be specified.

A cautious person. An optimistic person. A pessimistic person. A critical person. A creative person.

Even when facing the same problem, the response changes.

A cautious person searches for risks. An optimistic person searches for possibilities. A critical person searches for flaws. A creative person searches for new ideas.

The same thing happens in human society. Even when people attend the same meeting, they are not all thinking about the same thing. Personality changes what people notice.

A similar phenomenon appeared in AI.

Expert Prompting

A further development was Expert Prompting.

This is not about personality. It is about expertise.

For example, instruct the model: "You are an experienced cybersecurity expert."

The answer then begins to pay attention to security vulnerabilities. It considers attack paths. It considers access control. It performs risk analysis.

Ask the same question with a different instruction: "You are legal counsel."

Now the model looks at contractual issues. It looks at the scope of responsibility. It looks at regulatory risks.

In other words, an expert-like perspective appears.

The important point is that the AI has not truly become an expert. It has not acquired new knowledge. No additional training has taken place. What changed was the direction of the output.

But the result was still useful. Researchers saw the possibility of artificially creating behavior that resembled a team of specialists.

Role Prompting

These ideas eventually converged into Role Prompting.

Role Prompting is an extremely simple idea.

Assign a role to an agent. That is all.

You are an engineer. You are a product manager. You are a reviewer. You are a researcher.

Then, even if the model is the same, the output changes.

The engineer thinks about implementation. The manager thinks about priorities. The reviewer searches for problems. The researcher thinks in terms of hypotheses.

In human society, this is obvious. But in LLM research, it was an important discovery.

Because it meant that multiple perspectives could be drawn out of the same model.

Roles Do Not Increase Knowledge

There is an important point here.

Roles do not increase knowledge.

This is very easy to misunderstand.

Becoming an engineer role does not add new programming knowledge. Becoming a lawyer role does not mean the model graduated from law school. Becoming a doctor role does not mean the model studied medicine for six years.

The model's internal knowledge does not change. Its parameters do not change. Its training data does not change.

What changes is how knowledge is used.

And that difference turned out to be larger than expected.

Why Does Behavior Change?

A natural question arises here.

Why does behavior change merely because a role is assigned?

Only a few lines of prompt have been added. Why does the entire output change?

To understand this, we need to return to the Transformer itself.

The Direction of Attention Changes

As discussed in the Transformer article, attention lies at the center of LLMs.

Attention is the mechanism that determines what the model focuses on.

The model contains an enormous amount of knowledge internally. But it does not use all of it at once. For each question, it selects which knowledge to prioritize.

Role prompts influence this selection.

For example, if the model is told: "You are an engineer."

Then knowledge related to design, performance, maintainability, implementation methods, and technical constraints becomes more likely to be prioritized.

On the other hand, if the model is told: "You are a product manager."

Then market, customers, priorities, cost, and schedule come to the foreground.

The knowledge is the same. But the knowledge being retrieved changes.

That was the effect of roles.

The Same Thing Happens in Humans

In fact, this closely resembles what happens in humans.

Imagine an excellent engineer. Suppose that person becomes CEO the next day.

Their programming ability does not suddenly disappear. Their knowledge and experience remain.

But what they think about changes. They look at profit. They look at the market. They look at talent. They look at the competitive environment.

When a role changes, the direction of attention changes. As a result, decision-making changes as well.

What happens in AI is similar in structure. The knowledge did not change. The direction of attention changed.

Roles Are Cognitive Lenses

Researchers gradually began to understand this.

A role is not knowledge.

A role is a lens.

The same world is seen. The same problem is seen. The same information is seen. But when the lens changes, the appearance changes.

Engineers see technical problems. Legal staff see risks. Accountants see costs. Executives see markets.

Everyone is looking at the same reality. But they focus on different parts of it.

AI behaved in the same way. By assigning a role, researchers could change where the model looked. They could change what it valued. As a result, different behavior emerged.

Can Diversity Be Created Artificially?

Here researchers noticed an important possibility.

If roles can change the direction of attention, then even with the same model, it may be possible to create agents with different perspectives.

An engineer role. A researcher role. A manager role. A reviewer role.

All of them use the same model. But each looks at a different aspect of the problem.

If that is possible, the echo chamber problem might be reduced to some degree. Instead of mere copies of the same model, it might be possible to create a team with different roles.

A study that directly tested this hypothesis appeared in 2023.

That study was CAMEL.

The Emergence of CAMEL

By 2023, multi-agent research was beginning to accelerate rapidly.

AutoGen had demonstrated the possibility of AI-to-AI communication.

Role Prompting had shown that different behaviors could be elicited from the same model.

But one crucial question remained unanswered.

Can roles alone create a functioning team? Can roles alone generate meaningful diversity? Can roles alone improve problem-solving ability?

A study that confronted these questions directly was CAMEL.

CAMEL occupies a particularly important place in the history of agent research because it marked the transition from "AI agents that talk to each other" to "AI agents that collaborate through roles."

What Was CAMEL?

CAMEL stands for Communicative Agents for Mind Exploration of Large Scale Language Model Society.

The name is long. The core idea was remarkably simple.

Human behavior changes depending on roles. If that is true for humans, could it also be true for AI?

That was the hypothesis CAMEL set out to test.

At the time, many researchers were beginning to suspect that scaling models indefinitely was not the only path forward.

Larger models. More data. More parameters.

These were important. But human civilization was not built solely by increasing individual intelligence. It was built through cooperation. Through organizations. Through division of labor.

If those mechanisms were essential for human intelligence at scale, perhaps they could also become essential for AI.

CAMEL was one of the earliest attempts to explore that possibility.

AI User and AI Assistant

At the center of CAMEL were two agents.

An AI User. And an AI Assistant.

This was an unusual design.

In ChatGPT, the human is the user and the AI is the assistant. In CAMEL, both participants are AI.

The AI User presents a goal. The AI Assistant responds. A conversation unfolds.

The key idea was that the two agents possess different roles. They are not merely copies of one another. They have different responsibilities. Different perspectives. Different objectives.

Because of those differences, meaningful interaction becomes possible.

A Software Development Example

Consider a software development task.

The AI User acts as a product manager. The AI Assistant acts as a software engineer.

The product manager describes requirements. The engineer proposes implementation strategies. Questions are raised. Requirements are clarified. Specifications become increasingly concrete.

This is exactly the kind of conversation that occurs daily inside human organizations.

What makes it remarkable is that both participants may be powered by the very same language model. The only difference is the assigned role.

Inception Prompt

The true heart of CAMEL was something called the Inception Prompt.

Today, role prompting is common. At the time, it was surprisingly novel.

Each agent begins by receiving a definition of who they are.

For example:

You are an experienced software engineer. Your responsibility is to develop practical and implementable solutions. You prioritize technical constraints and engineering feasibility.

Meanwhile, another agent receives a different definition.

You are a product manager. Your responsibility is to maximize user value. You focus on requirements gathering and prioritization.

These instructions dramatically change the direction of the conversation.

Task Specifier

CAMEL included another important mechanism: the Task Specifier.

Researchers quickly realized that vague goals lead to vague discussions. When objectives are unclear, conversations drift.

The same thing happens in human organizations. Projects fail when nobody clearly understands the goal.

For this reason, CAMEL introduced a process for clarifying objectives before collaboration begins.

What are we building? What are we trying to achieve? What does success look like?

Only after those questions are answered does the interaction begin.

This simple step significantly improved the stability of the system.

Why Was the Task Specifier Important?

Many discussions of CAMEL focus on role prompting. But the Task Specifier was equally important.

Roles alone cannot solve everything.

A brilliant engineer cannot succeed without a clear objective. A skilled manager cannot rescue a project whose goals are undefined.

The same principle applies to agents.

Role and objective. Only when both are present can genuine collaboration emerge.

How Does the Conversation Progress?

CAMEL interactions follow a structured process.

A goal is presented. Roles are assigned. Agents begin discussing the task. Plans are proposed. Problems are identified. Solutions are revised. New proposals emerge.

The process looks remarkably similar to the workflow of a human team.

Researchers observed something fascinating. Agents with different roles genuinely approached problems from different directions.

The engineer focused on implementation. The product manager focused on user value.

The same model generated different perspectives. This was precisely what Role Theory predicted.

CAMEL Experiments

The researchers tested CAMEL across a wide variety of tasks.

Software development. Education. Creative projects. Planning tasks. Problem-solving exercises.

Again and again, agents with assigned roles naturally began dividing labor.

Managers organized objectives. Engineers designed implementations. Questions generated clarifications. Requirements evolved. Conversations progressed.

In some cases, solutions emerged that a single agent had failed to discover.

This suggested that role-based collaboration could genuinely improve outcomes.

Did AI Truly Understand Roles?

An important clarification is necessary.

CAMEL did not prove that AI truly understands roles. At least not in the human sense.

The engineer agent does not genuinely believe it is an engineer. The product manager agent does not think it works for a company.

What is happening internally is much simpler. The model generates language patterns associated with engineers. Or language patterns associated with managers. It reproduces those patterns.

Yet the practical result was surprisingly powerful. Role specialization emerged. Division of labor emerged. Organizational behavior began to emerge.

Resonance with Society of Mind

Many researchers reading CAMEL were reminded of Marvin Minsky's classic theory: Society of Mind.

In 1986, Minsky proposed a bold idea.

Intelligence is not a single unified thing. Instead, intelligence emerges from the cooperation of many smaller agents.

At the time, implementing such a system was extraordinarily difficult. Computational resources were limited. The idea remained largely theoretical.

The arrival of LLMs changed that.

CAMEL unintentionally recreated part of Society of Mind in a modern form. Instead of symbolic agents, language-based agents could now collaborate through conversation. The theory suddenly felt much more concrete.

What CAMEL Demonstrated

CAMEL did not demonstrate that AI possesses consciousness. It did not demonstrate that AI possesses identity. It did not demonstrate that AI possesses genuine expertise.

Its contribution was more practical.

Assigning roles can produce different behaviors from the same model. Those behavioral differences can improve collaboration. And collaboration can sometimes improve problem-solving ability.

This insight profoundly influenced later agent research.

Researchers began realizing that model size was not the only factor that mattered.

Organizational structure matters. Division of labor matters. Coordination mechanisms matter.

Agent research gradually expanded from studying individual intelligence to studying collaborative intelligence.

And at the same time, a new set of questions began to emerge.

If roles work so well, are roles alone enough?

CAMEL's Success and New Questions

CAMEL had a significant impact on agent research. The reason was simple.

Roles worked.

Earlier multi-agent research had focused primarily on the existence of multiple agents. CAMEL shifted attention toward something else.

The number of agents was not the most important factor. Roles were.

Even when agents were powered by the same model, assigning different roles produced different behaviors. Those behavioral differences created collaboration.

This discovery was deeply attractive to researchers because it suggested a path toward improved performance without endlessly increasing model size.

But almost immediately, new questions emerged.

Did CAMEL actually create specialists? Or were researchers merely observing AI systems that behaved like specialists?

Did AI Truly Become Specialized?

This remains an important question even today.

An engineer agent. A legal advisor agent. A researcher agent. A manager agent.

Are these truly different experts?

Strictly speaking, no.

The underlying model is usually identical. The knowledge is largely identical. The training data is identical. The parameters are identical.

In that sense, they are not genuine specialists.

Yet their behavior is undeniably different.

The engineer prioritizes implementation. The legal advisor prioritizes risk. The researcher prioritizes hypotheses. The manager prioritizes outcomes.

This resembles human specialization surprisingly well.

CAMEL did not create expertise itself. It created expert perspectives.

Role Drift

CAMEL also revealed limitations. One of the most important was Role Drift.

Role Drift refers to the tendency of agents to gradually move away from their assigned role over time.

An agent may initially behave like an engineer. But as a conversation grows longer, it may begin making managerial decisions. Or it may start commenting on areas outside its intended responsibility.

The same thing happens in human organizations. Meetings become longer. Discussions become more complicated. Responsibilities become blurred.

AI agents exhibited similar behavior. Assigning a role did not guarantee that the role would be maintained indefinitely.

Infinite Loops

Another well-known problem was the Infinite Loop.

Agent A proposes a solution. Agent B revises it. Agent A proposes another revision. Agent B revises it again. The process continues endlessly.

Human organizations often experience something similar. Meetings never end. Reviews never end. Requirements never stop changing.

Multi-agent systems encountered the same issue. Especially when goals were vague, conversations sometimes lost track of termination conditions.

Roles enriched discussions. But they could also prolong them indefinitely.

Coordination Failure

An even more difficult problem soon appeared: Coordination Failure.

As more roles are introduced, more agents are introduced. As more agents are introduced, coordination becomes harder.

Who makes the final decision? Who owns responsibility? Who resolves disagreements?

Role assignment alone cannot answer these questions.

Human organizations face exactly the same challenge. A collection of talented specialists does not automatically become an effective organization. Organizations require coordination mechanisms.

CAMEL demonstrated the value of roles. But it also revealed that organizational problems still remained unsolved.

What CAMEL Left Behind

Despite its limitations, CAMEL's contribution was enormous.

Because it changed the way researchers thought about agent systems.

Before CAMEL, the central question was: how can we make a single model smarter?

Larger models. More data. Longer context windows. Better reasoning.

After CAMEL, another possibility became visible.

Cooperation. Division of labor. Role specialization.

Perhaps intelligence could emerge not only from larger models, but also from better organization.

This idea would become one of the defining themes of later agent research.

A Curious Similarity to Human Society

Looking back, there is something fascinating about this story.

Researchers were attempting to build AI. Yet in many ways they ended up rediscovering human society.

Division of labor. Roles. Responsibility. Cooperation. Review. Decision-making.

These are not new inventions. Human societies have relied on them for centuries.

CAMEL was significant because it recreated part of those mechanisms inside AI systems. That is why the paper left such a strong impression on the research community.