Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

What, exactly, is an 'AI Agent'? Here's a litmus test

tines.com

Post reply on HN
Print view View original

11–20 of 45 posts

Re: What, exactly, is an 'AI Agent'? Here's a litmus test

#11

Post by wongarsu »

ninininino wrote: post #4
wongarsu wrote: post #3

Earlier quoted context omitted.

I agree with you. People really overcomplicate this. From wiktionary: """ Agent (plural agents) - One who exerts power, or has the power to act. - One who acts for, or in the place of, another (the principal), by that person's authority; someone entrusted to act on behalf of or in behalf of another, such as to transact business for them. - [various more specific definitions for real estate, biology, etc] From Latin a…

The problem with just using that definition is that drawing the line of what it means to have "the power to act" or to "act for, or in the place of, another" is subjective. Most would agree that a system or automation that could receive the instruction "do my entire job for me" and proceed to physically embody a bio-clone of me, walk to my office, impersonate me 40hrs a week, and keep my pay check coming in while I p…

It all comes down to your definition of "act". Which maybe does split into at least two criteria: the "trigger" (is running "git push" every time I ask it to enough, or does it have to decide to do that on its own, for example by monitoring my workflow) and the "action" (is running "git push" enough, or does it have to be able to order movie tickets?).

On the action my view is fairly lax. Anything that modifies the world counts, which does include a git push run on my computer. Tasks aren't less real just because they have a convenient command line interface.

The trigger is a bit trickier. We expect the agent to have some form of decision-making-process (or at least something that looks and feels like one, to avoid the usual discussion about LLMs). If a human doesn't make decisions they are a tool, not an agent. Same rule for AI agents. But defining the cut-off point here is indeed hard, and we will never agree on one. I'm not at all opposed to deciding that IFTTT is an agent, and that slapping some AI on it makes it an AI agent.

Re: What, exactly, is an 'AI Agent'? Here's a litmus test

#12

Post by rpgbr »

namaria wrote: post #8
whiddershins wrote: post #5

"AI Agent", as of April 2025, is a largely meaningless term that serves as a descriptive abstraction for people who aren't so familiar with coding. Debating its definition is pointless. AI Agent is whatever you want it to be if it helps you design or explain something.

I think there's a contradiction in claiming a term is meaningless and discussing its definition is pointless, and saying it can help explain things.

It's somewhat like “metaverse”, which to this day nobody can explain precisely, and yet we spent two years talking about it as it was the future. Meaningless and subject of countless discussions. There's no contradiction; it is possible.

Re: What, exactly, is an 'AI Agent'? Here's a litmus test

#13

Post by janalsncm »

It’s a bit amusing that so much ink has been spilled over what the definition of an “AI agent” is.

I don’t care. I care what your software can do. I don’t care if it’s called AI or machine learning or black magic. I care if it can accomplish a task reliably so that I don’t have to do it myself or pay someone to do it.

We had the same argument about 3 years ago when everyone started calling things “AI”. They use LLMs to generate text. Usually they have outsourced all of the interesting technical work to a handful of providers backed by big Web 2.0 companies.

Re: What, exactly, is an 'AI Agent'? Here's a litmus test

#15

Post by marxplank »

andy99 wrote: post #7

Anthropic has a definition: Workflows are systems where LLMs and tools are orchestrated through predefined code paths. Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks https://www.anthropic.com/engineering/building-effective-age... While I know it's a marketing term, I think a good distinction is that agents hav…

the LLM does have the ability to send garbage output in protest

Re: What, exactly, is an 'AI Agent'? Here's a litmus test

#16

Post by tedk-42 »

No such things yet. Just marketing hype for a product people are creating.

It's currently a blanket term for gluing together a series of interactions via code and relying on LLMs for interpreting input or creating output data.

LLMs, no matter how clever can go off right now an execute an API request (e.g. execute something in a bash terminal like `curl -XPOST --data 'blah' https://api-endpoint>`).

Re: What, exactly, is an 'AI Agent'? Here's a litmus test

#17

Post by ajcp »

> Our litmus test for AI agents: Does the AI system perform actions under its own identity?

So service accounts are agents? This seems pretty thin.

In AI an "agent" is simply any code/workflow/automation that utilizes an LLM to respond to broadly defined external/environmental stimuli and decide how to react given broadly defined motivations and/or objectives.

Not agent: Document comes in -> If invoice return key-value pairs to make API call with.

Agent: Document comes in -> You're a finance professional: decided what to do with this document. Here are the tools/actions available to you: X, Y, Z.

Both use AI and can achieve the same thing, but one is "agentic", while the other is deterministic.

Re: What, exactly, is an 'AI Agent'? Here's a litmus test

#18

Post by manojlds »

andy99 wrote: post #7

Anthropic has a definition: Workflows are systems where LLMs and tools are orchestrated through predefined code paths. Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks https://www.anthropic.com/engineering/building-effective-age... While I know it's a marketing term, I think a good distinction is that agents hav…

But they muddle it by saying all of them are Agentic Systems

Re: What, exactly, is an 'AI Agent'? Here's a litmus test

#19

Post by falcor84 »

janalsncm wrote: post #13

It’s a bit amusing that so much ink has been spilled over what the definition of an “AI agent” is. I don’t care. I care what your software can do. I don’t care if it’s called AI or machine learning or black magic. I care if it can accomplish a task reliably so that I don’t have to do it myself or pay someone to do it. We had the same argument about 3 years ago when everyone started calling things “AI”. They use LLMs…

It's fine that you as a user of these systems don't care, but nevertheless this is useful terminology for people looking to design such systems.

Re: What, exactly, is an 'AI Agent'? Here's a litmus test

#20

Post by float4 »

andy99 wrote: post #7

Anthropic has a definition: Workflows are systems where LLMs and tools are orchestrated through predefined code paths. Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks https://www.anthropic.com/engineering/building-effective-age... While I know it's a marketing term, I think a good distinction is that agents hav…

I should have read this 12h ago! This afternoon, I tried to create my first simple agent using LangChain. My aim was to repeatedly run a specific python analysis function and perform a binary search to find the optimal result, then compile the results into a markdown report and export it as a PDF.

However, I now realize that most of these steps don't require AI at all, let alone agents. I wrote the full algorithm (including the binary search!) in natural language for the LLM. And although it sometimes worked, the model often misunderstood and produced random errors out of the blue.

I now realize that this is not what agents are for. This problem didn't require any agentic behavior. It was just a fixed workflow, with one single AI step (generating a markdown report text).

Oh well, nothing wrong with learning the hard way.

Post reply on HN