In recent years, AI has become the new Klondike, sparking a gold rush that gains momentum with every new trend. One of the latest buzzwords in this AI frenzy is "AI-agent." But what does that even mean? A robotic James Bond?

New tech trends rarely live up to the hype, and the AI vocabulary is already messy and crowded with multiple misconceptions. Adding yet another term like "AI-agent" might seem unnecessary. After all, implementing technology for technology’s sake has often led to costly mistakes and poor returns. However, AI-agents might be different. While they may not come equipped with gadgets and a dry martini of a certain 007 - at least, not yet - they have the potential to deliver real, tangible value.

Motivated by the promise of AI-agents, this article aims to explore their value, practical use cases, how they work, and how they differ from the equally popular concept of AI-assistants.

I think, therefore I am… An AI-agent

While the famous quote from the French philosopher René Descartes might be a bit of a stretch here, it’s a fitting way to illustrate how AI-agents differ from AI-assistants. While an assistant can execute commands like “write a song,” it relies entirely on human input to know what task to perform. In contrast, an AI-agent can reason, make decisions, and autonomously execute actions without needing direct instructions for every step.

A tangible example - Johns Plant Business.

John runs a small business selling plants. A few years ago, he began collecting data on which plants sell best during different seasons. However, running the store leaves him little time to analyze the data. Fortunately, he has a trusty AI-assistant. Whenever John notices a drop in sales, he asks the assistant, “Which plants are in season?” Based on the assistant’s response, John decides what to order and manages to boost his sales.
While the assistant adds value to John’s business, its usefulness entirely depends on how well John utilizes it. John still makes all the critical decisions, such as which exact plants to buy, in what quantities and when.

Now imagine John had an AI-agent instead. The AI-agent is connected to John’s sales data and has access to the website where he orders his stock of plants. The AI-agent continuously monitors John’s inventory and the seasons. Based on historical data and current trends the agent reasons about which plants to order and in what quantities.
John is thrilled with how much work the AI-agent automates. However, he is also a bit nervous about giving too much control to the system. To address this concern “human-in-the-loop” functionality is added to the agent. This means that before the agent places any orders, John is notified and has the option to approve or reject the purchase.

To summarize:

• AI-assistant: Relies on John’s input to function - and the value is limited by how much input John provides.

• AI-agent: Uses data to autonomously reason and make decisions, offering value that is not dependent on John’s level of input.

AI-agent use-cases

Hopefully, the example of John’s plant business provided a clear distinction between AI-agents and AI-assistants. However, you might still wonder what exactly AI-agents can do?

Fortunately, AI-agents are already being implemented in various use-cases that highlight the specific pains they can address and the value they create. The following examples showcase AI-agents' potential in real-world applications from different domains.

Research & Development:

• Pain: The process of doing research and development requires several steps of manual work in searching previous articles and screening summaries. Often, this is based on static search queries with limited syntax opportunities. This leads to manual overhead and increased risk of errors.

• Value: AI-agents (and AI-assistants) can improve the process of research and development by improving the quality and time of the scientific information retrieval and screening process. Scientists would save time on creating search queries and save time on having to review fewer articles.

• Capability: Using AI-agents to support the screening and rating of scientific studies based on text inputs on inclusion and exclusion criteria.
  
  

Trading - Energy & Commodities:

• Pain: Navigating complex markets and placing orders manually can be time-consuming with multiple steps increasing the chance of errors. Especially when monitoring fluctuating prices.

• Value: AI-agents simplify the trading process by autonomously monitoring market conditions, identifying opportunities, and placing orders for specified amounts, all while keeping a human in the loop for approval.

• Capability: By combining market data analysis and natural language processing, AI-agents understand instructions such as “buy 10 options of product X at price Y” and execute the task efficiently. This led to faster and easier trading execution saving ~ 1 min on transactions while also lowering the error rate and risk of mistakes.  
  
  

Email search and efficiency:

• Pain: Most people in corporate business struggle with searching through thousands of emails for a specific keyword, time-stamp, or context. It is a tedious and inefficient process for most yet very common.

• Value: AI-agents drastically reduce search time by understanding abstract queries like “Find the email from my boss where he added inputs to my financial presentation” and pinpointing the exact email.

• Capability: Leveraging natural language understanding and contextual search, AI-agents can interpret and execute nuanced, human-like search commands. One of the key examples of this is https://superhuman.com/ with customers experiencing value of time saved and greater capacity.

Supply Chain - Demand Planning

• Pain: A common challenge is to handle the demand planning process in the front-line. This can be on the floor in retail stores, in warehouse management or in production planning. Across these scenarios a combination of historical demand and forecasting needs to be combined with orders, planning and logistics.

• Value: AI-agents can help ease and mature the process of demand planning by providing efficient and precise help in execution.

• Capability: Combining information from historical data and forecasting with workflows and external factors.

Financial Analysis:

• Pain: Finance professionals often do repetitive analysis and controlling manually comparing Excel sheets or multiple datasets. 

• Value: AI-agents can provide accelerated financial analysis and controlling tailored to the organization and domain requirements.

• Capability: The use-case of this can be both internal and external. In JP Morgan a part of the “LLM Suite” rollout includes the opportunity for employees to get internal support from AI-agents and AI-assistants on financial analysis.

Idea to code

• Pain: Incremental development with AI-assistants often requires constant manual intervention and lacks foresight for complex builds.

• Value: AI-agents can autonomously reason through the steps needed to build an application from an idea, reducing manual workload and speeding up development.

• Capability: By reasoning about the overall project, AI-agents can generate coherent, sequenced code that aligns with the intended application architecture.
  
  

E-commerce

• Pain: Customers often face hurdles like incomplete transactions, limited personalization, or lack of follow-up communication.

• Value: AI-agents enhance the e-commerce experience by placing orders, tracking shipping, facilitating image-based searches, sending cart abandonment reminders, and providing personalized recommendations based on previous purchases.

• Capability: Through integration with e-commerce platforms, AI-agents analyze customer behavior, access real-time data, and offer tailored interactions to drive sales and customer loyalty. Shopify are a great example of this (https://www.shopify.com/magic).

How does AI-agents work?

The next section of this article delves into the technical foundations that enable AI agents to operate autonomously. At the heart of what makes AI-agents unique is their ability to reason and take actions based on that reasoning. To understand what reasoning means for AI agents, it is essential to know the building blocks of AI-agents and understand the concept of control logic.

Building Blocks of AI-Agents

The foundational components of AI-agents work together to enable reasoning, decision-making, and action execution. These building blocks include:

1. Large Language Models (LLMs): LLMs serve as the cognitive engine in many AI-agents. Trained on vast amounts of text data, these models provide the ability to interpret language, generate coherent responses, and reason through complex tasks. LLMs act as the primary source of "knowledge" and general-purpose problem-solving capabilities for AI-agents.

2. Tools: Tools extend the capabilities of AI-agents beyond their language models. These include APIs, scripts, databases, or software applications that agents can invoke to perform specialized tasks, such as retrieving real-time information, analyzing data, or executing complex workflows. Tools are essential for integrating the agent's logic with the external environment.

3. Data: AI-agents rely on data as input during operation to make informed decisions. This can include structured data (like databases), unstructured data (like text or images), and real-time inputs from user interactions or external sensors. Access to relevant and up-to-date data allows AI-agents to adapt to changing environments.

4. General Software Components: These components provide the infrastructure and support systems necessary for AI-agents to function effectively. They include frameworks and libraries that handle memory, networking protocols, user interface elements, and integration with external systems. Software components are the foundation for AI-agents and ensure they interact seamlessly with their environment.

The synergy between LLM's, tools, data, and software components enable agents to interpret inputs, process information, and execute actions in a way that mimics human reasoning and behavior. By combining these components, AI-agents can bridge the gap between static, predefined programs and dynamic, intelligent systems capable of autonomous action.

Control logic

Control logic, a fundamental principle in engineering and computer science, refers to the set of rules, algorithms, or systems that dictate how a process or system behaves. It determines how inputs are interpreted, decisions are made, and actions are executed to achieve a desired outcome. In essence, control logic defines the path from input to output, with the elements along that path dictating what actions are performed between the two.

When applied to AI-agents, there are two main forms of control logic that are relevant: programmatic control logic and agent-driven reasoning:

• Programmatic control logic relies on fixed, rule-based instructions that dictate an agent’s behavior and ensure predictable outcomes within defined parameters. This approach is valuable in structured environments with clear and consistent rules, where consistency, precision, and strict adherence to guidelines are more important than adaptability or reasoning.
  
  

• Agent-driven reasoning by contrast, allows AI-agents to adapt dynamically and optimize outcomes in complex and uncertain situations. This reasoning enables AI-agents to determine the path between input and output, autonomously deciding which actions to take. It allows them to operate effectively in dynamic environments by leveraging available data and the learned logic from large language models (LLM's). As a result, AI-agents can navigate ambiguity and perform tasks that demand flexibility and critical thinking.

In many real-world scenarios, where conditions are dynamic and unpredictable, the rigidity of programmatic control logic becomes a significant limitation. In such cases, AI-agents excel due to their ability to reason and adapt, making them better suited to handle the complexity and fluctuations of real-world environments.

The technology stack around AI-agents

Understanding  AI-agents, AI-agent components and interfaces often require an overview of the technology stack involved. From the most common perceptions around Foundational Models to the AI-agent specific technologies. Below is a tech stack overview with the key conceptual tech stack layers with some of the most common technologies/vendors: