Agency vs. Autonomy: Untangling Two Core Ideas in Agentic AI

Recently, more attention has been going toward understanding Agentic AI, and digging into how AI agents work surfaces two terms that sound almost interchangeable at first glance:

Agency.

Autonomy.

On a first pass, these two concepts blur together. Agency gets defined as a system's capacity to act, make decisions, and pursue goals. Then autonomy shows up right after, described as the ability of an agent to figure out how to reach a goal using whatever information and resources it has, without a human walking it through every step.

That's where the confusion sets in. If agency already covers deciding and acting, and autonomy also involves deciding and acting, what's the point of having two separate labels for what looks like the same behavior?

That confusion turned into a useful learning moment once the concepts got untangled.

First, I Needed to Separate the Two

The way through was to restate each term in plainer language.

For agency, the question becomes:

Can the system decide and act toward a goal?

For autonomy, the question becomes:

How independently can the system make those decisions and carry them out without a human directing each move?

That's the missing piece. Agency is about whether the capability to act and decide exists at all. Autonomy is about how much freedom the system has when exercising that capability. Once framed this way, the two ideas stop overlapping and start describing different dimensions of the same system.

An Agent Can Have Agency Without Complete Autonomy

An example makes this concrete. Say an AI agent receives a goal:

Resolve a customer's payment issue.

The agent assesses the situation and reasons through it:

I should check the payment status.

Then:

I should retry the failed transaction.

Then:

I should notify the customer.

Here, the system is working out which actions make sense. That reasoning process is agency.

Now imagine the same agent has to pause before every single action and check in:

Agent:
"I want to check the payment system."

Human:
"Approved."

Then:

Agent:
"I want to retry the transaction."

Human:
"Approved."

The agent is still the one deciding what it wants to do next. But a human remains the gatekeeper who approves each step before it happens. So the agent has agency, yet its autonomy is tightly constrained.

Now Increase the Autonomy

Now give that same agent a goal instead of a checklist:

Goal:
Resolve the customer's failed payment.

And let it work through the process on its own:

Inspect payment status
        ↓
Identify the problem
        ↓
Choose the appropriate tool
        ↓
Retry the payment
        ↓
Verify the result
        ↓
Notify the customer

Notice what's missing: there's no instruction like

First do this.
Then do this.
Now call this API.
Now check this.
Now send this.

Only the goal was specified. The system figured out the sequence of actions on its own, drawing on the information and resources it had access to. That's the moment autonomy stops being an abstract definition and becomes something you can actually observe.

The real question isn't just:

Can this system make decisions?

It's also:

How much human oversight does it require while making and executing those decisions?

The Mental Model Worth Keeping

After wrestling with all these definitions, a refreshingly simple model emerges.

AGENCY

Can I decide and act toward a goal?

And:

AUTONOMY

How independently can I decide and act
without step-by-step human direction?

That's the whole thing.

There's no need to carry a lengthy definition around in your head.

Agency → ability.

Autonomy → independence.

Then Agentic AI Starts Looking Different

Before reaching this point, it's tempting to picture AI applications along a simple flow:

User asks something
        ↓
Model generates something
        ↓
User receives response

But agents add another layer entirely.

The system might carry a goal.

It might have awareness of its surroundings.

It might have tools and resources it can draw on.

And rather than being handed every intermediate step explicitly, it can work out which actions make sense on its own.

That shifts the picture of what an agent actually is:

GOAL
  ↓
Observe available information
  ↓
Decide what to do
  ↓
Use available resources
  ↓
Take action
  ↓
Continue toward the goal

Within that picture, agency and autonomy each take on a distinct job.

Agency is about whether the system can decide and act at all.

Autonomy is about how much independence it has while doing so.

Why This Distinction Matters for Engineers

Knowing a definition on its own rarely feels like enough. It helps to keep pushing further with questions like:

"Why does this term need to exist?" "How does it differ from the one just covered?" "What job does it actually do inside the system?"

The same pattern shows up with agency and autonomy.

Initially, the mental map might look like this:

Agency → decision-making
Autonomy → decision-making

???

Once you dig into what separates them, it becomes:

Agency
→ capability to decide and act

Autonomy
→ independence in deciding and acting

Now each concept occupies its own distinct slot in how these systems get modeled mentally.

That's far more useful than just memorizing two dictionary-style definitions.

The Takeaway

The biggest shift in understanding comes from realizing that a system doesn't become autonomous the moment it's capable of making a decision.

There are actually two separate questions worth asking.

Can the system make decisions and take action toward a goal at all?

That's agency.

And:

How much independence does it have while exercising that capability, without a human directing each move?

That's autonomy.

The gap between them is subtle.

But once it clicks, agentic AI stops feeling so abstract.

That's really the rewarding part of learning in this space.

A new term shows up.

It gets questioned.

It gets held up against what's already known.

It gets tested for weak points.

And eventually a model emerges that's simple enough to explain to someone else without needing to reference the source material.

That's usually the signal that something has actually stuck.

Agency tells you whether a system is capable of deciding and acting. Autonomy tells you how independently it can carry that out.