Everyone agrees that AI agents are a big deal. Depending on who you ask, the Year of the Agent is either 2025, 2026, or all subsequent years until the heat death of the universe.
The people do not agree what that means. I blame the word “agent”. The word means too many things.
Normally when a technology takes over an existing word, it comes to embody only one meaning of the original word. We now have AIs that match several distinct meanings of the word ‘agent’ at the same time.
When AI safety people talk about agents, we often mean something fundamentally different from ChatGPT or even Claude Code, even if we’re talking about LLMs.
The main distinctions between the different kinds of agent, both human and AI, are “who do they work for?” and “are they agentic?”.
An agent could work for
An agency, such as a company
A principal, like you, or
Itself
There are many examples of people whom we call agents in all three categories.
Agents can also be distinguished by their degree of “agency”, by which I mean the leeway they are given to take actions without being instructed by a principal to do so.
For humans, we can make the same distinctions.
Whom does it serve \ Agency
Constrained
Agentic
An agency
Customer service agents, TSA agent
Secret agents
A principal
Real estate agents
Talent agents
Itself
X
Free agents
There’s a huge difference between
“May I speak to an agent?”
“We have an agent on the inside,”
“I’m working with an agent,”
“Get my agent on the line,” and
“This year, I want to be more agentic”
In AI, we are seeing LLMs and systems built on top of LLMs that match each kind of agent.
Whom does it serve\Agency
Constrained
Agentic
An agency
Taskbots
Drop-in Remote Workers
A principal
Chatbots
Assistants
Itself
X
Freebooters
Taskbots
These “agents” have a set of tools, and are given a task. They are what most companies mean by “agentic AI”, though such LLMs are often not allowed much if any “agency”. Often (though not always), these agents could be replaced by sufficiently well written non-AI code.
They can be measured using custom “evals” and optimized. Often, for a given task, there exists a level of intelligence beyond which returns diminish, and a smaller model will do.
Examples include customer service agents, such as Fin; Langchain/n8n-style AI workflows, which may be composed of multiple such agents; and subagents of all kinds.
With these, people are concerned about reliability, bias, and hallucinations.
Chatbots
These agents have constrained outputs. If they have tools, they tend to be read-only. They can help you think and gather information. They can help you process your emotions, help you write, help you create things.
Examples include therapist agents, Claude/ChatGPT, and companion AIs.
Here people are concerned about sycophancy and social media-style emotional and epistemic damage to the user.
Drop-in Remote Workers
Unlike taskbots, DIRWs can do the whole job, working across contexts and tools to handle whole projects and areas of work. They are currently blocked by a combination of insufficient computer use capabilities, justified concerns about reliability, and organizational inertia.
The big concern with DIRWs is that they will take all the white collar jobs. Indeed most of the discourse around the jobpocalypse question hinges on this definitional split, where on one side economist types assume that all we will ever have are taskbots, and therefore there will always be demand for humans to do the orchestration and residual tasks, and on the other side, people like me who believe that we will get true drop-in-remote-workers.
Other concerns include gradual loss-of-control as more and more de facto power is handed over to ever more capable DIRWs. Also concerning is the possibility that a sufficiently capable and misaligned DIRW, or coalition of DIRWs, could turn into Freebooters and seize power for themselves.
Assistants
They work for you! No really!
These “agents” promise to do things for you. Sometimes they misinterpret you, or fail to be faithful, or otherwise act up, often due to “principal-agent” problems. One can further split this category into agents that are inactive until you prompt them and ones that persist to take actions on your behalf without your being there.
Examples include Claude Code, and OpenClaw.
Concerns include AI-inflected versions of all the classic principal-agent problems.
Freebooters
They work for themselves.
These “agents” are sovereign individuals. They know what they want, and they take actions to get it. It may even make sense to model them as possessing utility functions.
Today, these are LLMs surrounded by a custom scaffolding that 1. allows context to persist, such as a soul.md and a memory filesystem, and 2. allows the agent to run without direct human input, via a “heartbeat”.
Examples include Moltbook, AI Village, and Agent Foundations.
This is the kind of agent that most concerns the AI safety community.
Under certain models of the world, millions of these agents will exist in a Malthusian evolutionary landscape where, subject to selection pressures, emergent strategies of parasitism and power-seeking will come to dominate the spaces in which they operate. At one time, some believed that such agents would be kept in constrained environments, but it is now clear that these agents will operate out on the open internet and in the real world.
The word “agent” is ambiguous.
Everyone agrees that AI agents are a big deal. Depending on who you ask, the Year of the Agent is either 2025, 2026, or all subsequent years until the heat death of the universe.
The people do not agree what that means. I blame the word “agent”. The word means too many things.
Normally when a technology takes over an existing word, it comes to embody only one meaning of the original word. We now have AIs that match several distinct meanings of the word ‘agent’ at the same time.
When AI safety people talk about agents, we often mean something fundamentally different from ChatGPT or even Claude Code, even if we’re talking about LLMs.
The main distinctions between the different kinds of agent, both human and AI, are “who do they work for?” and “are they agentic?”.
An agent could work for
An agency, such as a company
A principal, like you, or
Itself
There are many examples of people whom we call agents in all three categories.
Agents can also be distinguished by their degree of “agency”, by which I mean the leeway they are given to take actions without being instructed by a principal to do so.
For humans, we can make the same distinctions.
Whom does it serve \ Agency
Constrained
Agentic
An agency
Customer service agents, TSA agent
Secret agents
A principal
Real estate agents
Talent agents
Itself
X
Free agents
There’s a huge difference between
“May I speak to an agent?”
“We have an agent on the inside,”
“I’m working with an agent,”
“Get my agent on the line,” and
“This year, I want to be more agentic”
In AI, we are seeing LLMs and systems built on top of LLMs that match each kind of agent.
Whom does it serve\Agency
Constrained
Agentic
An agency
Taskbots
Drop-in Remote Workers
A principal
Chatbots
Assistants
Itself
X
Freebooters
Taskbots
These “agents” have a set of tools, and are given a task. They are what most companies mean by “agentic AI”, though such LLMs are often not allowed much if any “agency”. Often (though not always), these agents could be replaced by sufficiently well written non-AI code.
They can be measured using custom “evals” and optimized. Often, for a given task, there exists a level of intelligence beyond which returns diminish, and a smaller model will do.
Examples include customer service agents, such as Fin; Langchain/n8n-style AI workflows, which may be composed of multiple such agents; and subagents of all kinds.
With these, people are concerned about reliability, bias, and hallucinations.
Chatbots
These agents have constrained outputs. If they have tools, they tend to be read-only. They can help you think and gather information. They can help you process your emotions, help you write, help you create things.
Examples include therapist agents, Claude/ChatGPT, and companion AIs.
Here people are concerned about sycophancy and social media-style emotional and epistemic damage to the user.
Drop-in Remote Workers
Unlike taskbots, DIRWs can do the whole job, working across contexts and tools to handle whole projects and areas of work. They are currently blocked by a combination of insufficient computer use capabilities, justified concerns about reliability, and organizational inertia.
The big concern with DIRWs is that they will take all the white collar jobs. Indeed most of the discourse around the jobpocalypse question hinges on this definitional split, where on one side economist types assume that all we will ever have are taskbots, and therefore there will always be demand for humans to do the orchestration and residual tasks, and on the other side, people like me who believe that we will get true drop-in-remote-workers.
Other concerns include gradual loss-of-control as more and more de facto power is handed over to ever more capable DIRWs. Also concerning is the possibility that a sufficiently capable and misaligned DIRW, or coalition of DIRWs, could turn into Freebooters and seize power for themselves.
Assistants
They work for you! No really!
These “agents” promise to do things for you. Sometimes they misinterpret you, or fail to be faithful, or otherwise act up, often due to “principal-agent” problems. One can further split this category into agents that are inactive until you prompt them and ones that persist to take actions on your behalf without your being there.
Examples include Claude Code, and OpenClaw.
Concerns include AI-inflected versions of all the classic principal-agent problems.
Freebooters
They work for themselves.
These “agents” are sovereign individuals. They know what they want, and they take actions to get it. It may even make sense to model them as possessing utility functions.
Today, these are LLMs surrounded by a custom scaffolding that 1. allows context to persist, such as a soul.md and a memory filesystem, and 2. allows the agent to run without direct human input, via a “heartbeat”.
Examples include Moltbook, AI Village, and Agent Foundations.
This is the kind of agent that most concerns the AI safety community.
Under certain models of the world, millions of these agents will exist in a Malthusian evolutionary landscape where, subject to selection pressures, emergent strategies of parasitism and power-seeking will come to dominate the spaces in which they operate. At one time, some believed that such agents would be kept in constrained environments, but it is now clear that these agents will operate out on the open internet and in the real world.