AI Agents. Making the Uncool - Cool Again

If you’re on the receiving end of the AI buzzword bingo - and we almost all are - AI Agents and agentic systems are now on the top. So, should you get excited about it? Should you be implementing it? Or maybe that is just a rebranding for the stuff you should be doing anyway? :) 

While most of the hype centers on splashy applications, let’s explore how AI Agents breathe life into the uncool use cases - the tasks nobody dreamt about but everyone relies on. 

From Generative AI to Practical Reasoning

Generative AI brought a lot of excitement because it was hard to crack and felt like alchemy. But once we wanted to move from parlor tricks like witty rap verses or funny stories to handling real use cases like customer support without supervision, we quickly discovered challenges and limitations. Hallucinations, an inability to perform math and reasoning accurately, or subtle errors when following longer instructions can derail even rudimentary use cases. It also creates this ambiguous feeling—how can something so smart be so dumb at the same time? Yet, it's all about appreciating the different kinds of abilities a model (or a system) can have and the different means of achieving them.

(BTW, I described vastly my approach to hallucinations in the previous issue of this newsletter.)

In many cases, reasoning might be easy if we have the right data in front of us and we know what rules or algorithms to follow. But without proper tools (or skills) and data, what is left might be more akin to a philosophical question than to decision-making—regardless of whether we have a human or AI in charge. No wonder the model is tempted to hallucinate—lacking proper data to base answers on, and pushed for an answer, I'd do the same.

What practical usefulness demands of us is much less glamorous than GenAI—correct data and well-laid-out logic or rules. With this in front of an LLM, the task becomes much simpler. Maybe, in a sense, LLMs are meant to be the bridge and user interface we all needed to bring AI to the masses. They serve as the intermediary that makes complex systems accessible, translating sophisticated backend processes into something we can easily interact with.

And if that's true, then we’re looking at a natural evolution: from LLMs to complex systems, where AI Agents can play a huge role. And, like always, when you strip the magic away, it means… more software.

What Are AI Agents?

Here’s a quick overview of AI agents to get on the same page:

AI agents are systems that integrate multiple components to deliver more effective results. At their core is a controller (usually an LLM)—the "brain" that makes decisions and coordinates with other key elements:

• Planner: Decomposes complex tasks into smaller, actionable steps. Planners can vary in flexibility, from adaptable LLM planners to completely predefined plans for greater reliability.

• Memory: Retains the history of previous steps and additional user or environment context, enhancing the agent's ability to make informed decisions.

• Tools: External APIs or functions that the agent can call to gather information or perform specific actions, such as submitting a form or querying a database.

In simple terms, AI agents perceive, reason, and act to achieve specific goals. For instance, an AI customer service agent that not only answers questions but also processes refunds or schedules appointments.

Think of it as a digital worker—not just responding to prompts but breaking down a problem, planning, and executing tasks with the right tools and data. Unlike standalone LLMs, AI agents are built to do, transforming AI from a passive conversational assistant into an active and capable executor.

For a deeper dive into AI agents, copilots, and chatbots, check out our detailed exploration: Why and How to Build AI Agents for LLM Applications

The Rise (and Rebirth) of AI Agents

The conversation about AI agents exploded over the last few months, bringing them to the forefront. Yet, the early wave of excitement hit much earlier with projects like AutoGPT, sparking visions of autonomous systems that could do it all: think, plan, and act, all on their own.

AutoGPT, for those who might have missed the fanfare, was a fascinating early attempt at making LLMs do more than answer questions. It aimed at something ambitious—taking goals and autonomously figuring out the steps to achieve them. The promise of AutoGPT felt like a next-level paradigm: you tell the AI what you want, and it thinks, plans, and executes. No hand-holding, just results. This is where the idea of "AI Agents" truly captured people's imagination—machines that could genuinely act with a purpose beyond simple prompts. The early attempts might have been too optimistic—trying to hop over difficulties with the sheer general capabilities of LLMs—but it seems we're finally finding our way now.

LLMs lean towards pattern recognition over understanding - they are good at educated guesses and simulate what we would consider thinking more than actually engaging in it. But they lack inherent logical consistency and struggle with precise reasoning or complex tasks without external help. For example, while an LLM might falter when on an analytical case interacting with a database and requiring exact calculations, an agentic system, equipped with specialized modules like tools and hardcoded APIs, provides the structure and guardrails needed for such work. This has shifted the focus from standalone LLMs to complex systems composed of smaller, specialized components—such as data connectors or integration layers—where predefined boundaries ensure reliability. It’s not flashy—but it works. And by “just getting it done,” AI agents bring more excitement into this mix.

AI Agents Bringing Back the Uncool Use Cases

AI agents are shining a light on use cases that weren’t exactly cool during the early generative AI hype—things like structured workflows, data extraction, integration tasks, and rule-based systems. These might have seemed dull compared to generating creative content, but they are incredibly useful.

Example #1: Database agent

General-purpose AI agents often fall short when interacting with databases, e.g., due to reliability issues. In contrast, specialized agents like a Text2SQL agent can interact with databases using well-defined actions, providing greater reliability, performance, and avoiding potentially dangerous SQL injections. This specialized approach ensures secure, consistent interactions that are essential for critical business applications.

Example #2: HR automation at scale

Global manufacturing company struggled with its HR platform’s complexity, dragged by legacy systems. Employees found it challenging to update basic information or perform basic tasks due to the difficulty of finding and following instructions across multiple systems. The AI agent automates such updates. First, by understanding the manual instructions and interpreting HTML and visual elements of websites, the agent uses an LLM to generate automation code (series of “clicks” and/or API calls), which is validated and tested. Then, when triggered, the agent executes the automation according to the request. By reducing user friction and improving process efficiency, this AI agent is designed to scale beyond HR, automating complex tasks across various software systems within the organization.

Example #3: AI Agents for Scalable Customer Support Automation

We developed an AI-driven conversational agent to automate appointment booking, reducing the reliance on human staff and extending the hours of operation. The agent used structured conversation workflows and integrated with the client’s scheduling system via APIs for real-time availability checks and confirmations. Supporting both text and voice interactions, it enabled patients to book appointments conveniently, including handling rescheduling and cancellations. Advanced natural language understanding ensured accurate responses, even in voice-based queries. This scalable solution provided 24/7 accessibility, improving response times while lowering operational costs and keeping the conversion rate. By streamlining the process, the agent allowed staff to focus on complex issues or on-site work. The overall patient experience was also enhanced.

Back to What Works and Back to the Future

In a sense, we’re returning to what works. When you need guarantees, defined error rates, and predictable outcomes, structured systems are simply more reliable compared to the flexible but unpredictable nature of LLMs. LLMs still have a place—they handle uncertainty, communicate with users naturally, and help guide the flow of processes. They act as both the user interface and, in some sense, the agent interface.

This means more software will be created at both ends of the spectrum—tools to get the most out of powerful models, as well as systems that act as a bridge between users, tools, and existing infrastructure. Think ChatGPT Canvas or Claude Artifacts on one end of the spectrum and just released Model Context Protocol on the other.   

This may be where AI agents truly find their niche: combining the best of both worlds. LLMs bring human-like interaction and general capability, while traditional software brings the reliability and precision we often need. And in that marriage, AI agents bring together the practical, the reliable, and the innovative—making the old new again, and the uncool cool again.

Robert