One of the most common uses of Generative AI we see in enterprise businesses focuses on making complex processes more simple, more agile and less costly. In exploring how to leverage Large Language Models (LLMs), Agent Systems have emerged as a new paradigm with their own set of challenges and opportunities.
For the uninitiated, Agent Systems utilise the LLMs’ apparent reasoning capability (at least within contexts it was trained in) and internal world models to retrieve relevant information and take actions based on it.
While this approach overcomes many limitations of early LLM applications - such as restricted information retrieval and limitation of only text-based interface - it also introduces new challenges. These include increased system complexity, higher fragility, and longer response times.
If you’re looking to automate complex processes within your organisation and considering an LLM-based Agent approach, you should know the strengths and weaknesses of the approach to conduct a thorough risk-benefit analysis tailored to the use case. We’ll provide you with an overview of the evolving LLM-based approaches and the key strengths and weaknesses of the Agent method.
The development of Large Language Models (LLMs) accelerated dramatically with the mainstream launch of ChatGPT, building on foundational work from previous decades. Early models struggled with any generalisation due to limited scalability. However, recent breakthroughs in model architecture have led to exponential growth in both sophistication and size (over four orders of magnitude increase in six years), making LLMs capable of handling a wide range of natural language tasks. This progress has transformed LLMs into key components of products like ChatGPT and Google’s Gemini.
Despite their capabilities, LLMs have inherent limitations that create challenges for enterprise applications. While they can process large amounts of text and provide reasonable responses, their internal world model is static—restricted to the knowledge available at the time of training—and quickly becomes outdated. Furthermore, LLMs lack memory and can only handle isolated interactions, making it difficult to manage multi-step workflows or maintain context over time. Crucially, LLMs are designed to generate responses, not necessarily to solve problems. They may provide incomplete or inaccurate answers without prompting for essential follow-up information. For example, when asking for a potential holiday destination, the model will return a wide range of options without first enquiring about budget or travel preferences.
To address these shortcomings, methods like Prompt Engineering and Retrieval-Augmented Generation (RAG) have emerged. Prompt Engineering helps guide the model’s reasoning by crafting specific instructions, while RAG allows the model to pull information from external databases to enhance accuracy.
While useful for targeted tasks, these techniques fall short when applied to more complex enterprise workflows, where decision-making based on up-to-date information and multi-step processes are essential. This is where Agent Systems come in—aimed at overcoming the limitations of LLMs by integrating tools, memory, and reasoning to deliver more robust, end-to-end solutions.
The LLM Agent System architecture expands on earlier techniques by integrating three key components: Tools, Memory, and Reasoning. Together, these enable LLMs to perform more complex, multi-step tasks.
The LLM’s primary role shifts from simply generating responses to orchestrating these components to meet user objectives, using a crafted prompt to guide the system's operations.
The Agent System architecture significantly enhances the LLM's ability to provide relevant, actionable information and perform tasks autonomously.
Unlike traditional automation methods such as Robotic Process Automation (RPA), which are rigid and fragile, LLM-based systems can adapt dynamically to new inputs without being explicitly reprogrammed.
The use of external Tools and memory storage also mitigates many of the shortcomings of LLMs, such as outdated information and limited capacity for exact calculations.
The Agent System architecture can improve the quality of the information that the LLM is using to make decisions and give it the ability to act (such as book a holiday package). The inherent flexibility of LLMs means that the developed automation is not as fragile compared to traditional methods such as Robotic Process Automation (RPA). It can complete tasks without being explicitly programmed.
The combination of Tools made available to the System can directly address LLMs’ limitations of its information relevance and ability to carry out exact calculations.
In the case of the holiday planning assistant, upon receiving our holiday preferences, it may take note of that in its Memory, then access a database of holiday destinations that match them. Following this, it may make an API call to a booking site to check for the availability and latest prices. Based on the returned list of options, the user then picks a specific holiday package, which the Agent System can book via a separate API call.
While Agent Systems offer clear advantages, they also inherit the limitations of LLM-based technologies, with added layers of complexity.
The increased number of LLM calls compounds the risk of errors. A misinterpretation of a retrieved piece of information can lead to a further mistake in the Reasoning system, sending the whole system down an undesired processing pathway. As the complexity of the solution increases, its error proneness and fragility increases while troubleshooting becomes more challenging.
To address these risks, various checks can be introduced. These range from explicit rules that enforce system constraints to additional LLM calls for validation.
However, each approach has trade-offs: rules can improve robustness but reduce flexibility, while extra validation steps may further complicate the system.
The increased number of calls also results in higher costs and latency. This impacts the ROI consideration of any use cases - especially the public-facing ones as the number of interactions may be more challenging to anticipate and manage.
Managing a single LLM-based component can be a challenge - for example, over-engineered prompts can hinder upgrades to a newer model and diagnostics of problems can be difficult. When multiple LLM-based components are used in an Agent System, these challenges multiply, making maintenance and scalability even more complex.
It is crucial to understand the transformative potential of Agent Systems and how they can streamline and automate complex processes within organisations. These systems can bring unprecedented flexibility and efficiency to enterprise operations, enabling automation that adapts dynamically. However, organisations must approach the adoption of this technology thoughtfully and strategically, remaining grounded in its current capabilities and limitations.
To maximise the chances of successful adoption in high-value applications, organisations should:
This iterative approach enables organisations to make better-informed decisions based on the actual capabilities of the technology and the value it delivers. The process will involve trade-offs—flexibility may need to be balanced against robustness, or speed against accuracy, depending on the specific application. By carefully navigating these trade-offs, enterprises can adopt a tailored approach that meets their unique requirements while mitigating associated risks.
Ultimately, adopting Agent Systems is not just about staying ahead of the curve—it’s about making strategic, informed choices that align with your organisation’s needs and long-term goals.
