Killing Flies with Cannonballs

Headline image source: HIPPO Education - https://soundcloud.com/hippo-education/wdidn-killing-a-fly-with-a-cannonball

Summary

Despite all the hype, “agents” don’t really hit the mark as I would have hoped, whether it be in the efficiency, environmental, or even “agency” department. We’ll explore definitions, talk about some examples, and ultimately land on where I’m currently feeling about the new agent wave of generative AI.

Intro

Recently, I have been seeing a lot more talk about needing an “AI agent” for this, or needing “agentic behavior” via a workflow for that. I’ve been playing around with these tools myself locally via n8n to help with creating menus for the week to buy groceries for,. Had a prospective client reach out because they needed an agent for… checks notes… people to chat with a knowledge base? Even after explaining that a chatbot could do that more than well, it was still like talking with a wall.

Anecdotally, I also saw Sabrina Ramonov, a prominent AI educator and influencer, show off how you can get an agentic workflow to order a pizza for you (and even recognize when the store is closed) which is super cool! I think a lot of people deeply enjoy the fact interactions like Tony Stark with JARVIS are starting to become more real by the day. The example of ordering pizza shows how sophisticated these tools are getting and the benefits of using natural language in the context of an AI workflows’s actions and the tools they’re provided. With all of this being said though, isn’t all of this a little…overkill?

That’s where I want to head today with this post. Yes, agentic behavior can be useful, but the more I utilize these tools in my own life, the more I’m left wondering whether the juice is worth the squeeze.

We’ll cover the difference between agents, ai workflows, and automation, the effects of utilizing agents from multiple perspectives, and other thoughts that make me a bit skeptical about replacing all my simple API requests with an LLM.

Let’s dive in!

How much of these tools are actually “Agents?”

You’ll notice I try to avoid saying the word “agent” outright in regards to many of these tools. I do so because honestly? I’m not sure everyone is using the word right.

I’d like to start by defining agents in the context of what we’re discussing today. A majority of the tools out on the market aren’t actually agents, but rather agentic workflows. This post by Alexandre Katjas provides some key insight into what I mean.

There is a key difference between automation, an AI workflow, and an AI agent in the context of generative AI:

Automation - Probably what the majority of people are used to. You have a system with an API you want to connect to that lets you perform a predefined list of actions in their space, like adding contacts, removing subscriptions, or editing a blog post. The system you are using also has these predefined tasks. Not every system can talk to each other natively, so you need some elbow grease to make two systems sing the same song!

AI Workflow - I find this area to be where chatbots reign. There can be logic around input / output, but sometimes there isn’t really logic. You have a document (or set of documents!) you want a model to read, then be able to answer about after. What the user can ask is a nebulous void. Or maybe a flow to extract a recipe for patatas bravas from the mess of ads all over the page. Workflows can also delve into agentic behavior, but most are lacking one key trait to be considered an “agent.”

AI Agents - The key word I love that Alexandre uses in this graphic is “autonomy.” Rather than having predefined tools or boolean logic, an agent is able to make rational, last minute decisions regarding the task they’re given. “Highly adaptive” is another fantastic word in the graphic. To me, this speaks to a higher level of task completion. What if I ask for something that the agent can’t provide?

Are the agents in the room with us right now?

The biggest thing I can say about whether a “tool” built with generative AI is an agent is that the definition is somewhat muddy. Depending on who you ask, and whether or not they have a vested interest in protecting AI company shareholder value, you may get a wildly different set of answers. As an example: Operator from OpenAI might be an agent, since it claims to be fully autonomous with a web browser. It’s marketed as such, so it must be an agent, right?

If I create my own n8n workflow, incorporate Gemini 1.5 Flash to devise action plans to execute, then hook up tools like https://community.n8n.io/t/new-community-node-automate-browser-tasks-in-n8n-with-robot-framework/62201 and browser nodes to perform tasks, does that make it an agent? Or is it a workflow performing agentic behavior? Maybe!

Both situations are tough to answer. Yes, Operator could be an agent. Yes, I could build my own agent with open source tools. I mentioned this briefly, but I fundamentally believe the difference lies between whether the tool in question has that autonomy and is highly adaptive. I also think a third key variable in this defines agentic behavior: context. In what context do these tools act as agents vs. agentic workflows?

In the context of a web browser, if I tell Operator to find me the best pizza place in Fort Collins, it will likely make up some criteria as to what the best pizza place is. Maybe it will get Google Reviews to make the “decision”? Yelp? That’s pretty much like an agent by definition. Changing the context, however, might yield different analysis on whether Operator acts as an agent in that environment.

The bigger question in this situation is: do we need an agent to find the best pizza place in Fort Collins?

Do we need all this firepower?

I started my own AI agent journey by working with local enthusiasts in my area to explore common use cases and see what resources were already available. An example of this came when I was introduced to Nate Herk’s Youtube channel: https://www.youtube.com/@nateherk. Plugging in the “Ultimate Assistant” set of workflows into n8n, I ended up with an amazing set of tools to handle things like emails, calendars, contacts, even contact creation! It only took me 3 hours to get the model to create an all-day event on the calendar exactly like I asked it to!

Three things come to mind about agentic workflows and my experiences thus far: efficiency, environmental concerns, and agency.

Is it more efficient than doing it yourself?

The more I use agentic workflows, the more it makes me feel like I’m reinventing the wheel compared to the alternative. Here are some examples:

• I can set up a flow for an agentic workflow to add a calendar event with my partner for coffee… or I can set up an event on my calendar with my partner for coffee in ~2-5 minutes?

• I can automate my emails to show me summaries of emails I actually want to read… or I can set up filtering in ~5 minutes, block spammers with a couple clicks, and read the emails I actually want to read?

• I can tell an agent to order pizza for me… or order the pizza myself within ~5 minutes, assuming everybody knows what they want?

• I can get Operator to do a search of the top 10 sources regarding the topic I’m interested in… or I can do a google search? If not google, a yahoo search? DuckDuckGo?

Many of these workflows are amazing proofs of concept of what’s possible, but they carry too much overhead in multiple capacities to make for a meaningful solution to the alternative. Plus, going back to Alexandre’s example, the flexibility in the context of what an agent can do may conflict with whatever deterministic output you’re trying to achieve.

Are we forgetting the environmental elephant?

Another consideration: if AI tools were already burning energy, water, and precious metals with single calls… what about adding agentic behavior?

We’ve already discussed how ChatGPT takes an insane amount of power to run on this blog, linked below.

Agents make multiple calls back and forth to perceive, reason, and adapt their plan as they receive new information. Amechanistic OODA loop, if you will. Given that, according to Forbes, ChatGPT consumes approx. 0.25 kilowatts (or on the high end of how much it takes to run a desktop computer for an hour) and a plastic water bottle’s worth of water for a single request… agentic workflows blow those numbers out of the water!

Keep in mind, this would probably be significantly reduced if you were to use your own workflows locally. How many consumers would know how to install Docker, Ollama, and troubleshoot connecting these tools in order to save on energy though?

Who has the agency here?

The biggest, and albeit hardest to quantify, reason that these agentic workflows don’t sit right to me is the human aspect. What are you giving up by letting these tools think for you? Is it really that more convenient to let Operator do the work for you, to let Claude tell you what to say, and to let the slew of other models and products on the market guide you into what they want to do?

We already know these tools are biased. They’re made by humans, so of course they are. I’ve also written some on that based on a study from the University of Navarre:

With these biases in mind, I don’t believe using agentic workflows democratizes anything for you unless you have full control of all the pieces that make this happen. Especially if the model you’re using is in the cloud provided by one of the AI giants. Who’s to say ChatGPT can’t already give you sources based on their “industry partners” for what is the best product in a given market?

In many ways, I find the name of agents to be somewhat ironic in the context of what it’s doing to the average person. Many are pawning off their own agency to make an AI agent happen. Studies have started to come out regarding cognitive decline and dependency on AI tools. If AI is ordering your pizza, reading your mail, scheduling your meetings and doing everything you were doing in your life, where does that leave you? Personally, I feel like many people glossed over the message in WALL-E and decided it would be a great blueprint for our future.

Conclusion

All of this to say, consider where you’re implementing agents and whether these processes actually benefit from adding this behavior in. Choose the best tool for the job given budget, requirements, and other factors, not just the shiny new tool right in front of you. Right now? I think I will stick to ordering pizza myself.

This is where I make a novel comparison to killing a fly with a cannonball, but I think I’ve beat said fly enough in regards to this topic. Take care, and onto the next!

“Father, I’ve learned Jedi mind tricks. I can convince you to be my agent now to give me more cheeeeeese.”