September 12, 2024

AI Agents: AI Agents at Work

In our first post on AI agents, we examined the foundations of this technology, defined and explained what it does, and highlighted a series of real-world use cases across various disparate domains. In doing so, we alluded to their high-impact and utility nature, demonstrating how their versatility, sophistication, and customizability dynamically enable broad and narrow application initiatives throughout fundamentally different workflows. AI agents are beginning to redefine the generative AI (GenAI) frontier, and while it’s arguably still too early to tell to what extent, it’s clear that their impacts on the future of work will be profound and widespread.

To reiterate and build upon the discussion in our previous post, AI agents possess several properties that make them especially attractive in business settings. While no single one of these properties is individually unique to AI agents, a combination of them is. Each of these properties is explained below:

  • Agency → The ability to act in an environment to produce consequential effects on the environment or other entities (e.g., AI agents, humans, or other systems) within it. Agency and autonomy are closely related but distinct—autonomy primarily concerns the capacity for self-governance. In this respect, AI agents are autonomous and agentic—autonomy allows them to make choices and decisions without human intervention and agency allows them to act on these very choices and decisions, usually with some degree of human guidance or input.
  • Goal-Orientation and Complex Problem Solving → Given their foundation in GenAI, notably Large-Language-Models (LLMs), AI agents are designed to dynamically interact, via natural language, with human operators in complex environments. Through natural language prompting, users can provide detailed contextual explanations of certain problems, goals, or tasks, outline parameters or instructions for AI agents to follow, describe certain metrics or thresholds to adhere to, and intervene to modify how tasks, problems, or goals are accomplished. Since advanced AI agents also learn from and adapt to their environments and interactions with users, their goal orientation and complex problem-solving abilities should improve over time.
  • Customizability and Extensibility → Broadly speaking, AI agents can take the form of generalists or specialists, depending on what kind of post-training enhancements and fine-tuning is performed. Custom AI agents designed for targeted use cases like personal tutors, AI researchers, and customer service chatbots, can deliver results that are closely aligned with organizational objectives, values, and work processes while enabling human personnel to automate mundane or time-consuming components of their workflow. Conversely, generalist AI agents can interpret data across multiple domains and perform numerous tasks in technical and non-technical application contexts, supporting department-agnostic AI integration initiatives that are perhaps more easily scalable than their counterparts. In terms of extensibility, AI agents, being a GenAI technology, may also display emergent properties or capabilities as they scale, learn from, and adapt to their environments and users. Alternatively, some AI platforms—OpenAI and Mistral AI—also offer users the ability to build custom AI agents with unique capabilities repertoires without requiring fine-tuning.
  • Accessibility and Augmentation → Despite their sophisticated nature, AI agents are operated through conversational interfaces via natural language, making them accessible to non-technical users who typically don’t possess the skills required to leverage advanced computational technologies in professional settings. Similarly, AI agents also enable technical and non-technical users to utilize them for tasks that fall outside of their expertise—a software developer might use an AI agent to generate a written product description whereas a marketing specialist could use it to uncover data-driven insights into their marketing campaigns. AI agents are augmentative tools, and the more users experiment with them, the more they’ll be able to widen and enhance their existing skill sets.
  • Strategic and Long-Term Planning → Frontier AI models like ChatGPT, Gemini, and Mistral AI, despite representing the state-of-the-art, are still limited by their ability to generalize beyond training data to navigate novel, complex, or changing environments. Custom AI agents are more equipped to circumvent this problem, particularly when fine-tuned for specific operational use cases, objectives, or workflow processes. In such cases, an AI agent doesn’t have to generically anticipate how a given problem, situation, or task might unfold, especially in light of detailed user inputs, instructions, and proprietary training data. When custom AI agents are leveraged for strategic and long-term planning in targeted operational contexts, uncertainty will remain an important factor to consider, but the actions the agent takes are more likely to reflect organizational and user preferences contextually and accurately.

The core properties of AI agents make them attractive and lucrative tools. According to a recent survey, roughly 80% of tech executives indicate plans to integrate agentic AI systems into their organizations within the next 3 years. Moreover, when considering AI agents’ abilities to handle multiplicity in linear and non-linear workflows, easily interact with users of all profiles in natural language, and seamlessly integrate with and/or operate existing components of digital infrastructures, their status as high-value AI assets is deeply reinforced. AI agents are truly pushing the boundaries of what’s achievable with GenAI applications.

Throughout the remainder of this post, we’ll take a closer look at AI agents, beginning by discussing the different forms they can take and offering a broad functional overview of what working with an AI agent might look like. Following this, we’ll dive into a series of potential real-world AI agent use cases, breaking down each in detail to create an actionable impression of AI agents at work. We hope that by closely examining this topic, we’ll provide readers with the insights they need to begin exploring and developing their own agentic AI solutions within concrete operational contexts.

AI Agents: Form and Function

While the vast majority of AI agents are built on similar foundations, the term “AI agent” is still imprecise, failing to capture the nuanced forms this technology can assume. For instance, ChatGPT, Claude, and Perplexity are all examples of LLMs, but to say that they are the same would be inaccurate—ChatGPT is a multi-modal generalist chatbot with an expansive capabilities repertoire, Claude is a proficient writing assistant with impressive reasoning skills, and Perplexity is an AI-powered search engine fine-tuned for academic and scientific research.

Just as understanding the different forms of LLMs is crucial for comprehending the capabilities and limitations of certain Frontier AI models, the same applies to AI agents. In this respect, we articulate the various forms AI agents can take below, drawing from Amazon’s AI agent classifications while also including some of our own.

  • Learning Agents: Arguably the most advanced singular AI agent, learning agents learn and adapt to their previous experiences and operating environment to continually optimize the results they deliver. These agents can iteratively improve themselves by designing and implementing novel training tasks while leveraging certain input and feedback data streams to adapt to their environments dynamically. For example, an AI tutor that gathers real-time learning insights and data from individual students to personalize lesson plans.

  • Hierarchical Agents: A collective of AI agents structured hierarchically whereby each agent is organized in tiers, with lower-level agents performing individual tasks that are the result of complex task deconstructions performed by higher-level agents. Each agent acts independently while corresponding with the agent immediately above and below it, while the agents in the top tier ensure that all actions taken within the larger system are coordinated in the interest of a common goal.

  • Utility-Based Agents: AI agents that act as utility maximizers on behalf of users. Such AI agents possess moderately advanced reasoning capabilities and work toward achieving the outcome that is most rewarding for the user. To arrive at a desired outcome, the agent evaluates potential outcomes in terms of utility values, or perceived user benefits. An AI agent that integrates with a user’s calendar and email to schedule and summarize meetings, add important events to their calendar, and generate automated email responses and notifications is one such example.

  • Goal-Based Agents: AI agents that are prescribed specific goals and/or rules to follow when they’re tasked with achieving a desired outcome. Such agents are efficiency maximizers, tending to select the most direct goal-oriented path, and display higher-level reasoning capabilities, evidenced by their ability to compare and contrast complex actions in terms of their expected results. Most state-of-the-art general-purpose AI models like ChatGPT, particularly when they’re fine-tuned for a specific task repertoire, can assume the form of goal-based AI agents.

  • Simple Reflex Agents: Rudimentary AI agents that leverage data at their immediate disposal and are bound by specific functional parameters that restrict their ability to generalize beyond the tasks they are trained to perform. Such agents are useful in simple tasks like automated FAQs and password resets.

  • Model-Based Reflex Agents: Like their less sophisticated simple reflex counterparts, model-based agents are fairly similar, with the main difference being their decision-making mechanism. Such agents evaluate several decision paths and their respective outcomes before orchestrating a decision, and usually via additional data, can develop an internal world model that is eventually leveraged to drive and support certain decisions. Algorithmic trading agents, for instance.

  • Personalized Agents: AI agents that are designed or customized by a single individual or organization to perform one or several tasks directly relevant to a specific workflow. Such agents are typically either built on open-source pre-trained GenAI models or within Frontier AI platforms that offer the opportunity for custom AI agent creation. These agents will most often leverage state-of-the-art foundation models, and take the form of personal assistants that directly integrate with certain sub-processes in a given workflow.

  • Multi-Agent System: Hierarchical agents are multi-agent systems, but not all multi-agent systems are hierarchical. Some multi-agent systems might contain a series of AI agents tasked with distinctly different objectives—each agent is driven by its own optimization function, maximizing the efficiency with which a set of pre-defined objectives is reached. Alternatively, other systems may function in a semi-hierarchical or sequential manner, whereby only certain actions taken by agents are coordinated in the interest of a common goal, and overseen by agents in the top tier. Finally, multi-agent systems can also involve both humans and AI agents, working together in separate teams or as a whole to complete various tasks within digital or physical environments. It’s worth noting that multi-agent systems have yet to be tried and tested in real-world professional settings.

  • Recursive Self-Improvement Agents: Like learning agents, recursive self-improvement agents can also design novel training tasks for themselves to expand their capabilities repertoire or patch model vulnerabilities. However, these agents are further distinguished by their ability to modify and improve their source code autonomously to alter their optimization functions and decision-making parameters—advanced future versions might even independently self-replicate or create novel agentic AI systems. It remains to be seen whether such agents will ever become commercially available since the risks they inspire are predominantly existential, especially in loss-of-control scenarios.

  • Adversarial Agents: AI agents exclusively designed to probe, stress-test, or identify vulnerabilities in other AI models and agents, or alternatively, cybersecurity frameworks. These agents will likely be used in tandem with human specialists to streamline adversarial testing and red-teaming procedures for high-risk and/or high-impact AI models. In the future, these agents could also prove highly useful in risk assessment procedures for critical digital infrastructures composed of numerous complex and interconnected digital components and sub-systems, each with their own array of potential failure modes.

Now that we’ve covered the forms AI agents can take, we can break down, in simple terms, what a generalized interaction with an AI agent would look like in most cases. We describe the interaction sequentially, examining each step from a user perspective.

  1. User Query → The user formulates a simple or complex input query (i.e., prompt), usually in natural language, describing the task, objective, or question they have in mind, the parameters they want the model to follow in generating an output, and any examples of what a desirable output would look like. With models that offer speech capabilities, user queries can be administered in spoken language.

  2. Further Clarification → If more detail or information is required, the model probes the user to expand upon their input query, highlighting the components of the prompt that require further clarification. This process may suffice in one or multiple iterative steps depending on the desired output of the user and the sophistication of the AI agent—how well it internalizes and understands the information contained in the user query.

  3. Processing → The AI agent processes the information in the user query, accessing and synthesizing relevant information from its knowledge base or other external data sources to produce a non-generic output.

  4. Output Generation & Delivery → The AI agent generates a contextually accurate response intended to satisfy the user’s preferences as stated in the initial query. The output can take several forms, including a discrete yet detailed answer to a question posed by the user, a description of the actions the agent will take to achieve a stated objective, or the means by which a given task will be completed, particularly if agent must recruit or control other technologies or devices to perform the task.

  5. Dynamic Adjustment → Where AI agents allow users to dynamically adjust and refine their prompts in real-time during output generation, users may find that before an output is finalized, further prompt refinement is required to hone in on a desirable output. When users decide to dynamically adjust prompts, the AI agent will pause its output generation until the initial prompt is finalized. This step is optional and depends on whether an AI agent offers this capability.

  6. Follow-Up → The AI agent asks the user whether they’re satisfied with the final output and, if relevant, provides follow-up questions or feedback that enable the user to further refine or specialize their interaction with the agent. If the user is satisfied with the final output, they can approve it.

  7. Task or Objective Execution → In cases where the user prompts the AI agent with a specific task or objective, all that’s left to do once the final output is approved is reaching the stated objective or performing the intended task. At this stage, no further user intervention is required, unless it’s revealed, through human oversight, that the processes, methodologies, or tools the AI agent recruits are somehow incorrect, fallacious, or misleading.

  8. Evaluation & Conclusion → Once the AI agent completes the objective or task it has been assigned, the user should evaluate whether or not it has been successful. To do so, the user must have a concrete idea of what a successful task or objective completion would look like, and metrics or thresholds against which to compare the AI agent’s success.

These steps, as stated above, broadly describe what a potential user could expect when interacting with an AI agent. Evidently, the nature of these interactions could shift or vary with respect to what kind of AI agent the user leverages and for what purpose. However, we strongly recommend that in all human-AI interaction scenarios, especially those involving agentic AI systems, users verify and validate AI outputs, and if relevant, scrutinize the procedures through which certain tasks or objectives are achieved.

AI Agents: Potential Use Cases

In this final section, we’ll explore three potential AI agent use cases across health and wellness, environmental conservation, and education. We’ve selected these use case domains to illustrate the versatility of agentic AI systems and in the hopes of going one step further, demonstrating their status as high-utility domain-agnostic technologies.

Case 1: Health and Wellness Smart Virtues Inc., a company that specializes in developing smart technologies for personalized health and wellness interventions, deploys its latest AI agent, NutriCoach, a model designed to leverage wearable data in real-time to provide user-specific actionable health and wellness lifestyle interventions. To access NutriCoach, users purchase and download the app directly onto their smartphones and pair it with their wearable device.

NutriCoach performs several functions, including autonomously interpreting wearable data like sleep patterns, heart rate variability, blood oxygen saturation, and VO2 max to generate daily lifestyle recommendations for users. Each of these recommendations appears as a push notification on the user’s smartphone—when the user clicks on the notification, they’re redirected to the application interface where NutriCoach breaks down the recommendation in terms of actionable steps the user can take. NutriCoach also offers users the ability to dynamically interact with these recommendations, allowing them to indicate certain preferences they may have in natural language, like dietary restrictions or pre-defined fitness goals.

For users interested in further customizing their experience, NutriCoach offers a “user profile” section where users can input, in detail, their epigenetic history, pre-existing conditions, previous health interventions, dietary habits, exercise regimens, and other relevant information. NutriCoach then leverages this profile information in tandem with information extracted from numerous medical databases to create user-specific health and wellness goals while also periodically alerting users when correlations between data points indicate an increased risk for certain medical conditions like heart disease or diabetes.

If users realize they need to book an appointment with a human specialist to resolve a personal health issue, NutriCoach identifies all local healthcare providers within the user’s proximity. Then, NutriCoach prompts the user to input their insurance information and appointment availability to locate providers who meet their immediate needs. Once a user has completed these steps and selected a healthcare provider, NutriCoach books an appointment on their behalf and adds it to their calendar.

Case 2: Environmental Conservation In an attempt to preserve biodiversity, the California-based environmental tech start-up, EcoResponse, builds Savior, an AI agent that can interpret data from a network of interconnected sensors and drones that work together to create an interactive real-time map of local ecosystems. In fact, Savior is also capable of autonomously controlling these sensors and drones and leverages advanced image recognition and sound analysis techniques to identify and track local fauna, detect illegal deforestation and poaching activities, and monitor fluctuations in vegetative growth.

The drones within the Savior network are all equipped with a series of sophisticated technological extensions. When poaching or illegal deforestation is identified, Savior triggers a built-in drone siren as a deterrent for further illegal activity while also leveraging facial recognition to identify poachers and deforesters, after which all relevant information is directly transferred to the nearest local law enforcement entity. Savior then relays this information back to EcoResponse’s proprietary database, summarizes it in natural language, and generates a report, which it then forwards to local law enforcement and the company’s key teams and personnel for further in-depth evaluation.

In cases where invasive insects destabilize an ecosystem, or where natural and man-made phenomena have resulted in diminished vegetative growth, Savior recruits a drone swarm, whereby each drone is equipped either with a local predatory insect species or seeds, which are then distributed selectively throughout affected ecosystems. Savior also leverages drone swarms to monitor migration patterns, feeding habits, and breeding among local fauna, transmitting this data to California’s Environmental Protection Agency (EPA), along with insights gathered from its other ecosystem sensors, to provide a list of environmental preservation recommendations.

Once the EPA receives these recommendations, it can accept, deny, or modify them as it sees fit—where recommendations are denied or modified, Savior prompts EPA personnel to explain the reasons behind their choices, after which Savior updates its parameters accordingly. Similarly, all ecosystem maps generated by Savior are made available via a publicly accessible database, allowing users of all profiles, from state rangers and farmers to hikers and EPA agents, to identify and communicate, via a simple conversational interface, where maps are outdated or inaccurate. In more nuanced cases, where individuals leverage Savior’s “visual ecosystem progression” feature, they may identify certain ecosystem trends that Savior overlooked—these trends can also be communicated via the same conversational interface.

Case 3: Education Education For You, a promising late-stage start-up created by leading school psychologists from around the world, releases EmpathAI, an AI agent that functions as a personalized emotional intelligence (EI) tutor that helps students build and cultivate crucial interpersonal skills during the most important stages of childhood development. Via simulated social interactions in virtual environments, EmpathAI leverages computer vision and NLP methods to interpret individual students’ speech patterns, vocal inclinations, emotional disposition or mood, and facial expressions.

To address students’ developmental needs, EmpathAI generates scenarios that test critical EI skills like conflict resolution, social and self-awareness, emotional regulation and empathy, achievement orientation, adaptability, and teamwork. However, to ensure that these scenarios reflect each student’s individual needs and age-appropriate standards in developmental psychology, EmpathAI begins by simulating a role-playing interaction with every student. Before this interaction starts, the student is prompted to create a personal avatar with a set of physical attributes that reflect the quality of “friendliness” in addition to another avatar, intended to represent EmpathAI, that reflects the student’s image of a friend.

Initial role-playing interactions with EmpathAI occur in virtual reality environments—to facilitate complete immersion—in which students’ avatars directly interact with their version of the EmpathAI avatar. The interaction takes approximately 20 minutes and is modeled after a game the student likes to play (e.g., tic-tac-toe, four-in-a-row, etc.). During the interaction, EmpathAI administers several increasingly targeted questions, interpreting the student’s speech patterns, vocal inclinations, mood, and facial expressions in real-time, and modifying each subsequent question or chain of thought with respect to the insights it gathers. Each question posed by empath AI is intended to probe both the emotional intelligence and maturity of the given student, and once the interaction concludes, an output score, evaluated across key EI metrics, is produced in natural language and made available to the relevant educator and parent.

If both parents and educators agree that further EI intervention is required, they can prompt EmpathAI to design a personalized lesson plan based on the results of the student’s initial assessment—EmpathAI then outputs a semi-structured lesson plan that leverages techniques like guided reflection, empathetic listening, and interactive storytelling to enhance the EI skills that a student lacks. Importantly, parents and educators can work with EmpathAI, as they would with an LLM, to iteratively refine this lesson plan until they’re satisfied, leveraging the agent’s ability to scan leading psychology databases for peer-reviewed studies, interpret and summarize university lectures, and offer specific insights on successful and unsuccessful EI interventions conducted in clinical settings. Once a lesson plan is finalized and implemented, EmpathAI autonomously generates biweekly notifications outlining individual students’ progress and reminding parents and educators that they can halt further interaction with EmpathAI at any time, particularly if the student begins exhibiting signs of insecure attachment.

Conclusion

Admittedly, there is a lot to digest in this post. We’ve covered the core properties of AI agents, the different forms they can take, the way they function in generic settings, and three in-depth potential AI agent use cases. However, we urge readers not to fret if they feel as though they haven’t grasped the full nature of this technology—AI agents are still very much in the early stages of their evolution, and while AI technologies do progress extremely fast, there’s still some time to explore what this technology is and can do.

If we were to offer one piece of advice to readers who wish to expand their knowledge of AI agents meaningfully and pragmatically, it would be this: experiment with as many different models as frequently as you can in a wide variety of use cases and environments. For readers who are rightly asking themselves, “What about the risks linked to AI agents? How can I experiment responsibly if I don’t know what to expect?” we answer: stay tuned, our next piece in this series will examine precisely this issue.

For readers interested in exploring other related topics in the AI landscape, innovation, and regulatory ecosystem, we suggest following Lumenova AI’s blog, where you can sift through more content on GenAI, responsible AI (RAI), and AI risk management, governance, and regulation.

For those who have already begun developing and implementing AI risk management and/or governance initiatives, protocols, or frameworks, we invite you to check out Lumenova AI’s RAI platform and book a product demo today.


AI Agents Series

AI Agents: Introduction to Agentic AI Models

AI Agents: AI Agents at Work

AI Agents: Potential Risks

AI Agents: Future Evolution


Related topics: AI Agents Generative AI Automation

Make your AI ethical, transparent, and compliant - with Lumenova AI

Book your demo