What is agentic AI? An introductory guide

Since the release of OpenAI’s GPT-3 large language model in 2020, LLM usage has grown steadily throughout the tech world. Since then, a bevy of AI firms, including Anthropic, Google, Microsoft, and Apple, have themselves released new models.

Those models are focused on solving different sets of problems, and each release looks to improve on both the reasoning capabilities and operational efficiency of previous models.

In 2024, leading models from Anthropic, OpenAI, and Google started pushing the forefront of what we know now as “agentic AI.” These generative AI tools offer a step forward in the working capabilities of LLMs of all stripes.

In this post, we’ll take a walk through the concept of agentic AI, learn how it works, and how you can leverage it to boost your own skills.

Definition and core concepts

To start with, let’s outline some of the basic technologies involved.

Large language model

A large language model, or LLM, is a software system that’s capable of interfacing through natural human language. Instead of relying on traditional programming languages for user interface, these programs interface through languages like English.

Large language models (LLMs) are one category within generative AI that allows users to specify what they’d like to generate via natural language, then examine the results that are generated.

Generative AI

Generative AI refers to a type of deep learning software model that can generate text, images, or videos.

More specifically, large language models (LLMs) are one category within generative AI that allows users to specify what they’d like to generate via natural language, then examine the results that are generated.

AI training

AI training involves feeding a very large amount of information into a specialized computer program to teach the statistical weight of correlation between different words, images, and videos.

Often, AI training involves multiple specialized steps of human curation as well as human feedback on the quality of responses during a testing phase. The majority of the cost involved with operating an AI model comes from the training phase.

AI model

The “model” in LLM refers to a combination of two things: a machine learning model combined with post-training programming to improve interactive behaviors. We referenced a model earlier: GPT3. You’ve probably seen discussion of other models like Claude 4, GPT5, or Gemini 3.

How agentic AI differs from other AI approaches

The basic idea behind an agentic LLM versus a more traditional LLM is that it provides the LLM with an interface to operate on your local computer.

This allows the LLM to do things like create, read, and edit files, run command-line commands, or collate information between local files and the internet. Most importantly, the agent has the ability to understand success and failure states based on user input.

This is what takes the AI from merely being an LLM interface to being a true “agent.” You can dispatch the agentic LLM to perform tasks, and it will iterate on the changes needed to achieve those tasks until it reaches a success state.

This kind of approach has proven especially fruitful for software developers. Because software has a number of easily automated success states—like automated tests, linters, and compilers—it’s easier for an agent to know whether it has completed a task successfully or not.

Real-world applications of agentic AI

The most prolific use of agentic AI today is in writing, debugging, and testing software. But that isn’t the only use case for agentic LLMs. Many users have adopted a workflow where they leverage an agentic LLM to organize notetaking apps like Obsidian.

As a heavy user of Claude Code, I like to think of it as a natural-language user interface for my PC. Anything that I can do by editing a file, invoking a script, or running a command-line command, I can do via Claude, and with an intuitive English-first interface.

Agentic AI in software testing and quality

Writing for Meta, Mark Harman explains it pretty clearly: “The rise of agentic software development means code is being written, reviewed, and shipped faster than ever before across the entire industry. It also means that testing frameworks need to evolve for this rapidly changing landscape. Faster development demands faster testing that can catch bugs as they land in a codebase, without requiring regular updates and maintenance.”

Let’s take a look at how agentic AI is paving the way for software teams to improve their testing today.

Faster development demands faster testing that can catch bugs as they land in a codebase, without requiring regular updates and maintenance.

1. Agentic testing

Agentic AI is useful for testing software in just about every imaginable context. We’re going to cover many of those in depth in subsequent sections, but it’s worth saying explicitly: if you already write software tests in any way, agentic AI will help you do it faster and often with higher quality.

Agentic AI will help you write unit tests, functional tests, regression tests, and performance tests. Oftentimes, it will do that by leveraging your existing testing tools.

However, you should challenge yourself to think of creative approaches to developing new test plans using agentic AI.

Tests that would have been prohibitively difficult or time-consuming to write in the past are accessible in a world where agentic AI will help you write the code. Let’s look at some of the different ways that’s true.

2. Agentic testing tools

In most contexts, the set of agentic testing tools looks exactly like the set of automated testing tools you already use.

If you’re already writing unit tests and running them as part of your CI process, then your best testing tool is simple: keep doing that. If you’re already building effective regression tests when you encounter a bug, leverage your agent to do that better and faster.

While agentic LLMs are powerful tools, they don’t mean that you need to throw everything you know out the window. You’re still focusing on writing smart tests with proven tools.

3. Agentic functional testing

Agentic AI is particularly helpful at one of the most challenging parts of functional testing: building out the context necessary for testing big, complicated features.

One of the hardest parts of writing good functional tests is that you often need to do a lot of test setup to make that work.

Agentic AI makes this much simpler. A well-fed agent is able to look at the context of the feature that you’re trying to test, and you can use natural English to describe the scenario that you’re trying to test.

Then, you can rely on the agent to generate the harness data that you need for running your functional test.

At a more advanced level, you might even rely on the agent to generate code which itself generates functional testing data for your tests.

One of the most substantial benefits of agentic AI is that writing code for smaller, less valuable tasks is much cheaper. AI agents are very effective at this kind of scripting that wouldn’t be worth doing if you had a human being doing it.

A well-fed agent is able to look at the context of the feature that you’re trying to test, and you can use natural English to describe the scenario that you’re trying to test.

4. Agentic regression testing

Agentic AI is also great at automating regression testing. The most critical part of regression testing is consistency. Regression tests are there to ensure that your code changes don’t break existing functionality within your app.

One of the more crucial factors in extracting value from regression tests is ensuring that every time you experience a bug, you write a regression test for that bug.

By this point, you probably already have a clear picture: agentic AI is very good at generating good regression tests. What’s more, it does it much, much faster than a comparable human regression test.

It’s trivial, after finding the root cause of a bug, to prompt your agent to generate a new test that encapsulates the issue. When you’ve fixed the bug, the test is always there, ensuring that you haven’t shipped a regression.

5. Agentic performance testing

Just like functional and regression testing, agentic AI is a huge help with performance testing. We covered this a little bit when we talked about functional testing, but agentic AI is even more valuable when working with performance testing.

One of the most difficult parts of working with performance testing is that we need to generate large amounts of load for however our application works. Doing that work manually is highly time-consuming. Doing that work with an AI agent? It takes a fraction of the time.

Agentic AI provides flexibility to develop a much wider variety of performance tests that quickly and simply fit into your existing performance testing regime.

And much like we talked about with functional testing, agentic AI makes it easier than ever to generate intermediate code that can be used to support your ultimate test harness.

6. Agentic test management

Agentic AI is a powerful tool for managing your test suites. One of the benefits of the agent’s natural language interface is that it very quickly acclimates to new text-based interfaces with no input from the user. Any discoverable command-line application is quickly “learned” by the agent.

This makes testing libraries a natural interaction point for AI agents.

As we noted previously, one of the real benefits of working with agents and software is that the agent is able to manage testing library outputs without any special training. The agent is able to read the output of test runs and identify which tests are failing without any human interaction.

At a more advanced level, this kind of interaction can lead to automated flows where you can point your agent at a failing test suite and ask it to identify common points of failure between the tests.

It’s also common to point agents at failing tests and ask them to either suggest fixes to the code to fix the tests, or automatically fix the tests itself.

One of the real benefits of working with agents and software is that the agent is able to manage testing library outputs without any special training.

7. Agentic quality assurance

Agentic AI is not itself a full replacement for human quality assurance, but it can do a great job helping to support QA engineers.

Also, agentic AI works well as a “pair programmer” for discussing different ideas. For a QA engineer, this is a real opportunity to explore ideas around how to most effectively ensure quality within the application.

Additionally, AI agents are able to interact with programs on the host computer through the browser. The agent’s ability to navigate web pages means that you can provide a set of English instructions for how to interact with the website, and then evaluate whether the flow produces the desired result.

In contexts like this, an agent can be a great addition to a QA engineer’s toolbelt.

Challenges, risks, and considerations

Working with AI agents is powerful. They’re terrific tools. But when you do, there are two key considerations to keep in mind at all times:

• How do I know that the agent did things the right way?

• How do I know that I’m working on the right thing with my agent?

Neither of these is a question that an LLM can answer for you. LLMs are great at automating certain kinds of work, and as we’ve noted, they have great testing capabilities.

But they’re not magic. They won’t simply allow you to eliminate things like critical thinking or domain knowledge.

Those questions are why we stressed the testing loop earlier in this post. Agents aren’t guaranteed to do high-quality work when you dispatch them on a task.

Hallucinations are still an ongoing concern. But testing loops allow you to constrain the agent and continue to point it in the right direction even when it tries to go off track.

Best practices

1. Define clear outcomes

The absolute best piece of advice anyone can receive while working with agentic AI is that the most important part of the entire process is to have a clear picture of what a successful outcome looks like.

The further you leave behind a clear picture of success, the more likely you are to fall into the pit of “AI slop.” Working with AI and providing minimal or low-skill oversight runs the risk of shipping low-quality code.

But when you use agentic AI to support your own skills, you can do great work more quickly than ever before.

The further you leave behind a clear picture of success, the more likely you are to fall into the pit of “AI slop.”

2. Develop a solid workflow

Beyond that, developers around the world have started to converge on what I’d call a “Research, Annotate, Plan, Implement” flow. Agentic AI is very good at processing natural language text, and it’s terrific at both making and following checklists.

By adapting your workflow to the flow most comfortable for the agent, you’ll improve your odds of getting good work the first time.

How Tricentis supports agentic AI–driven quality

Tricentis is leading the way in AI-driven quality for software teams in three different streams:

• Agentic test automation
• MCP servers for interacting with Tricentis services from your own AI agents
• A natural-language interface for orchestrating AI agents across the entire testing workflow

As we noted earlier, one of the major benefits of AI agents is that they improve the quality of the tools that you already use. Tricentis has taken this same approach with its own best-in-class tooling. Tricentis testing is already easy and powerful, and its AI agent tooling takes things to a whole new level.

This post was written by Eric Boersma. Eric is a software developer and development manager who’s done everything from IT security in pharmaceuticals to writing intelligence software for the US government to building international development teams for non-profits. He loves to talk about the things he’s learned along the way, and he enjoys listening to and learning from others as well.