Why Agents Are Becoming Harder to Classify
In the past six months, even many industry insiders have started losing their grip on what "Agent" means.
Yesterday it was Chatbot.
Today it's Coding Agent.
Tomorrow a General Agent appears.
The day after, a Vertical Agent.
Taxonomies, classification charts, four-quadrant frameworks — it's dizzying.
But I increasingly suspect the problem isn't that agents are too complex.
It's that we're looking at them wrong.
Most people classify by function.
Chatbots chat.
Coding agents write code.
General agents book flights and send emails.
Vertical agents know law or medicine.
This approach isn't wrong, per se.
But it explains less and less of what's happening today.
Because those boundaries are dissolving.
ChatGPT writes code.
Codex manages projects.
Claude runs workflows.
Vertical agents acquire general capabilities.
General agents keep absorbing domain knowledge.
And suddenly we realize:
These aren't different species.
They're more like different developmental stages of the same thing.
I recently revisited the history of agent evolution and noticed something that was hiding in plain sight: there have always been two paths.
The first is reasoning.
The second is workflow.
And everything happening in the agent space today is, at its core, these two paths converging.
Start with the reasoning path.
What made the earliest large models so striking?
Not that they could recall facts.
But that they could think.
Especially Chain of Thought — CoT.
Facing a complex problem, it reasons step by step.
Analyze.
Decompose.
Plan.
Arrive at an answer.
This is a purely cognitive trajectory.
The model increasingly resembles a thinking person.
Meanwhile, there's a completely different path.
The workflow path.
This one is far older than large models.
Older even than the internet.
Because every organization depends on workflows.
Companies run on them.
Governments run on them.
Software development runs on them.
Factories run on them.
Humanity's method for managing complex affairs is, at its essence, the SOP.
Break big tasks into small ones.
Define the steps.
Define the sequence.
Define the responsibilities.
Define exception handling.
Decades of automation have all belonged to this path.
RPA.
Scripts.
Assembly lines.
Automated approvals.
Automated deployments.
CI/CD pipelines.
All of it, at bottom, is workflow.
The difference is simply that processes were designed by humans.
And executed by machines.
So for a long time, the two paths ran in parallel, never touching.
AI handled thinking.
Workflows handled execution.
One was a brain.
The other a conveyor belt.
The truly interesting thing only started happening in the last two years.
Reasoning began reaching toward workflow.
Workflow began reaching toward reasoning.
At first, CoT was just a derivation process inside the model's head.
Then it became Planning — it started laying out plans.
Then Task Decomposition — breaking down tasks.
Then the Agent Loop — continuously revising plans based on environmental feedback.
And finally, today's dynamic workflows.
The other side was changing too.
SOPs used to be written by humans.
Flowcharts were drawn by humans.
Rules were set by humans.
Machines merely followed instructions.
Now we're seeing natural-language workflows.
Humans no longer specify every step.
They describe the goal.
The model generates the process.
Revises the process.
Decides the next move on its own.
And so we arrive at a genuinely important historical moment.
The two paths have converged.
Reasoning is no longer just thinking.
It has become action.
Workflow is no longer just rules.
It has acquired the capacity to reason.
Many people think of agents as an upgraded Chatbot.
That may not be the right framing.
From a historical perspective,
Agents look more like the marriage of CoT and SOP.
A fusion of reasoning systems and workflow systems.
Suddenly many phenomena snap into focus.
Why did Coding Agents mature first?
Because software development has always been a natural workflow.
Read the code.
Modify the code.
Run the tests.
Read the errors.
Modify again.
The feedback loop is crystal clear.
So reasoning and workflow fused here with the least friction.
Why have General Agents progressed so fast in the last two years?
Because at their core, they're trying to intelligentize every workflow in an open world.
Look things up.
Write documents.
Call tools.
Operate web pages.
Manage projects.
All workflow.
Why are Vertical Agents merging with General Agents?
Because domain knowledge, in the end, is just knowledge.
Law.
Medicine.
Finance.
Eventually it all comes down to task planning, tool invocation, and process execution.
The underlying architecture is converging.
So what we're seeing today is not that agent classification is proliferating.
Quite the opposite.
Different paths are flowing into the same river.
Chatbot.
Coding Agent.
General Agent.
Vertical Agent.
They look like they come from different worlds.
But they're actually heading toward the same destination.
That destination may not be a new product category.
It may be a new form of intelligent organization.
Once, humans designed processes and machines executed them.
Then, humans designed goals and machines generated processes.
Eventually, perhaps even the processes themselves will become dynamically evolved artifacts.
If the Chatbot era answered the question "Can AI think?",
Then the Agent era is really about answering:
How does AI turn thought into action?
And that, perhaps, is the most important — and most easily overlooked — thread running through the agent revolution of the last two years.