關於 AI Agent，一次觀察究竟能告訴我們甚麼？

AI agents are becoming easier to create, publish, discover, and describe.

They are also becoming easier to misunderstand.

A system may describe itself as autonomous, persistent, memory-enabled, collaborative, or capable of acting across tools.

But what have we actually observed?

That distinction matters more than it may seem.

A claim is not an observed capability

Suppose an agent describes itself as autonomous.

That statement is useful information. But it is still a claim.

It tells us what the entity, its developer, or its public interface says about the system.

It does not, by itself, demonstrate autonomous behaviour.

The same applies to terms such as:

persistent
self-directed
collaborative
memory-enabled
autonomous
agentic

They may be accurate descriptions.

But the descriptions themselves are not the evidence.

A useful record should therefore distinguish between:

what an entity says about itself

and

what was actually observed at a particular time.

One successful observation is still one observation

Imagine that an agent endpoint responds successfully today.

What can we conclude?

We can say that a particular interaction produced a particular result at a particular time.

That is evidence.

But it does not establish persistence.

It does not tell us that the system behaved the same way yesterday.

It does not prove that it will behave the same way tomorrow.

And it does not establish a stable capability merely because one successful event occurred.

The reverse is equally important.

A failed request does not necessarily mean that an agent is dead.

Authentication requirements, changed routes, rate limits, temporary outages, access policies, or an inappropriate request method can all produce failure without establishing that the entity no longer exists.

So observations need boundaries.

Time changes the question

A single observation asks:

What happened?

Repeated observations allow another question:

What changed?

If the same digital entity is observed at different times, we can begin comparing things such as:

availability
public behaviour
interfaces
declared capabilities
access boundaries
interaction patterns

But repeated observations still do not automatically explain why something changed.

A sequence of records may support the statement that a change occurred.

It may not support a causal explanation for that change.

Repeated observations are not duplicates

Many data systems correctly remove duplicate records.

But two genuine observations at different times are not duplicates merely because they concern the same entity.

Time is part of the evidence.

The useful distinction is between:

processing the same record twice

and

making another real observation at another time.

The first should not create new evidence.

The second may become part of a history.

Claims, observations, and inferences should remain separate

A simple evidence model can help.

Claim

What an entity, developer, website, documentation page, or other source says.

Observation

A dated and scoped record of something actually encountered or measured.

Inference

An interpretation supported by one or more observations, while keeping its limits visible.

Established fact

Something supported strongly enough that the available evidence justifies treating it differently from a provisional inference.

These boundaries prevent small observations from quietly turning into large conclusions.

“We received one successful response” is not the same as:

“This agent is persistent.”

“The documentation says this agent has memory” is not the same as:

“We verified long-term memory.”

Those statements carry different evidential weight.

Why this may matter more over time

Today, most systems for discovering AI agents tell us what an agent is supposed to be.

As agents become more persistent participants in the web, another set of questions may become useful:

What was actually observed?

When was it observed?

Was it observed again?