Workflows VS Agents

If you are using the Internet in 2025, it is hard not to hear about AI. And you’ve probably heard the term AI Agents as well, given they are currently very popular. But the question is, what actually are AI agents?

Definition of AI Agents

According to Google, AI agents are software systems that use AI to pursue goals and complete tasks on behalf of users. They show reasoning, planning, and memory and have a level of autonomy to make decisions, learn, and adapt.

But how do they do that? How can a system reason and plan your vacation for example? Well, it was a process to get them to where they are today. And it all started with prompting and LLMs. Now you might be wondering, what are LLMs?

Large Language Models (LLMs) are a category of foundation models trained on immense amounts of data making them capable of understanding and generating natural language and other types of content to perform a wide range of tasks.

(IBM)

And so, the first step to the evolution of AI agents was writing a good prompt which communicated with LLM, giving you the response based on the data it trained on. It was a good first step, but it had its flaws. One of the main issues was hallucination - if there was no answer inside of the knowledge base, it confidently made up an answer. Also, they lacked autonomy, meaning it couldn’t do stuff for you like book a flight because it didn’t have access to your data, nor was it able to produce information relevant to today’s date - its knowledge cutoff was the training date.

In order to fight the knowledge cut off and hallucinations, RAG came into play. So now we have yet another 3 letter abbreviation that we need to explain.

Retrieval-Augmented Generation (RAG) is the process of optimizing the output of a large language model, so it references an authoritative knowledge base outside of its training data sources before generating a response.

(AWS)

So how did it help the previous system? Well, before answering your question, the RAG part would retrieve relevant information and then, bundled with the model it provided you an answer. Given that you could get the relevant and more trustworthy information from it, this was a big step forward. However, at this point, this system was still lacking the autonomy - it could read the data but it could not act upon it.

And the next big step is when tools were added to this equation. Now with an LLM, RAG, and tools, you had a system that could not only use relevant, real-world information, but also the means to act upon that data.

Here is a real-life example of this. Imagine you want to find out about an event in the city and the schedule for a ferry to a nearby island. Your prompt would look something like this:

"Check the public event calendar for the city and find me a cultural event happening this weekend. Also, see what the ferry schedule to the nearby island looks like for Saturday."

The RAG pattern would be used to search for relevant events. The system would use a retrieval tool to connect to the city’s events API and get a list, for example:

• Event: Summer Nights - Outdoor Concert, Location: City Square, Date: Saturday
• Event: New Exhibition Opening, Location: City Art Gallery, Date: Saturday
• Event: Farmer's Market, Location: Central Market, Date: Saturday/Sunday

Now the "brain" of the operation, the LLM, will determine that out of this list, the "New Exhibition Opening" at the City Art Gallery best fits the "cultural event" criteria. But for the other part of your prompt (finding the ferry schedule) it’d need to call a specific tool. So, it would use a Public_Transit_API tool provided to it. Once it gets the schedule data from the tool (e.g., Departures: 08:00, 14:30, 19:00), it integrates this specific result into its final response.

Now, it combines all the data and provides you with an output:

"This Saturday, there is an opening for a new exhibition at the City Art Gallery. For travel to the nearby island on Saturday, the ferry from the main city port is scheduled to depart at 8:00 AM, 2:30 PM, and 7:00 PM."

When you compare it to the first step with only an LLM and a prompt, this is an enormous change in the quality of the results. So, does this combination = an AI agent?

Well, not exactly. This powerful system is still lacking a key ingredient: autonomy. You got the answer you needed, but it was just text. What makes an AI agent an agent is that it can use these components to autonomously take actions to achieve a goal. For example, after providing the answer, it could ask, "Would you like me to add a reminder for the exhibition and a calendar event for the 2:30 PM ferry to your calendar?" If you say yes, it then uses its tools to perform those actions for you.

So, if put on the spectrum of automation, AI agents would be on the one side. On the other, there would be workflows.

Why use workflows?

You might be wondering, with AI agents being so advanced, why would anyone turn to their opposite? Sometimes, there are cases where you want the whole process to be consistent and repeatable. Workflow can be described as a predefined, structured sequence of tasks designed to achieve a specific business outcome.

There are three main points of a workflow: the trigger, the actions and the logic. The trigger is the switch that starts the workflow. It is an event like, when a new email arrives, or a certain time and date occurs. The actions are the steps which the workflow performs. For example, email sending or adding an event to your calendar. And when it comes to logic you can think of it as paths that need to be taken, conditions that need to be met. If there is a word urgent in an email, notify the customer support, or for each new file do a certain thing.

So by combining a normal workflow with LLM, you get a powerful tool that can take off load from your employees by delegating repetitive automation tasks to AI.

One of the real world examples would be a candidate selection. If your HR team gets a lot of CVs for a new position, it would take them days to go through each one and read everything. So, the first process of eliminating the inadequate candidates could be delegated to the workflow. For example, if you are looking for a Senior Python Developer, you would make a workflow that would only consider people with Python mentioned with their skills, and probably 5 or more years of experience.

First action in that process would be data extraction. With a proper tool, you’d get extracted data from all the CVs. Then on the next step, you would prompt the LLM to do what needs to be done. For example, you could write something like:

"You are an expert HR assistant. Analyze this resume against our job description for a 'Senior Python Developer.' Return a JSON object with the following fields: 'years_python_experience' (integer), 'has_aws_experience' (true/false), 'mentions_fastapi' (true/false), and 'education_level' ('Bachelors', 'Masters', etc.)."

This is where the logic part comes into the play. These things need to be defined, for example, if years_python_experience is greater than or equal to 5 AND has_aws_experience is true… And with that being done, you would have a list of candidates that match the minimum criteria that could be reviewed among HR employees. The workflow could also do the part of rejecting ineligible candidates by sending them a generic message which thanks them for their interest in the company.

When to use AI agents and when to rely on workflow?

It really all depends on the use case and it is up to you to decide. The examples we have provided are just a few of many that happen in real life. If you are unsure about which approach to take, here is a table to take into consideration when you are making this decision.

With these technologies advancing every day and more models coming up, more and more menial tasks can be delegated securely to AI. We strongly advise you to think about saving your time and helping your employees by implementing workflows or AI agents to the adequate areas of your work.