“Meet the brain behind modern AI — the agent that acts.” “From automation to autonomy: The rise of AI agents.” “Learn what makes AI agents smarter than ever.”

What are AI Agents?

An artificial intelligence (AI) agent is a system or software that can do activities autonomously on behalf of either a user or another system by developing a workflow and using existing tools.

Beyond natural language processing, AI agents may do a variety of tasks such as decision-making, problem-solving, communicating with external surroundings, and carrying out activities.

These agents may be used to tackle complicated tasks in a variety of corporate scenarios, including software design, IT automation, code-generation tools, and conversational assistants. They employ advanced natural language processing techniques from large language models (LLMs) to understand and respond to user inputs step by step, as well as select whether to use other tools.

How does AI agent works

Large language models (LLMs) serve as the foundation for AI agents. For this reason, AI agents are frequently referred to as LLM agents. Traditional LLMs, like IBM® Granite™ models, rely on training data and have limited knowledge and reasoning capabilities. Agentic technology, on the other hand, employs backend tool calls to gather current information, improve processes, and construct subtasks autonomously in order to fulfill complicated goals.

Throughout this process, the autonomous agent gradually learns to conform to user expectations. The agent’s capacity to store previous contacts in memory and anticipate future actions promotes a more customized experience and thorough replies. This tool calling can be accomplished without human interaction, expanding the potential for real-world applications of these AI systems. Agents are defined by the following three steps or agentic components:

Goal initiation and planning

Although AI agents are independent in their decision-making processes, they require goals and specified rules established by humans. There are three primary factors on autonomous agent behavior:

• The development team responsible for designing and training the agentic AI system.
• The team that deploys the agent and grants the user access to it.
• The user assigns particular goals to the AI agent and determines which tools are accessible for usage.

Given the user’s objectives and the agent’s available tools, the AI agent conducts job decomposition to increase performance. Essentially, the agent devises a strategy including particular tasks and sub tasks to achieve the complicated goal.

Simple chores do not require planning. Instead, an agent may continuously reflect on and refine its replies without predetermining its next moves.

Reasoning with accessible tools.

AI agents take actions based on the information they perceive. Frequently, AI agents lack the whole knowledge base required to tackle all subtasks within a complicated objective. To address this, AI agents employ their available tools. These tools may include external data sets, online searches, APIs, and even other agents. After obtaining the missing information from these tools, the agent may update its knowledge base and engage in agentic reasoning. This implies that at each stage, the agent reassesses its strategy and self-corrects, allowing for more informed decision-making.

To assist demonstrate this process, consider a user who is organizing their trip. The customer asks an AI agent to determine which week of the next year would have the finest weather for their surfing trip in Greece. Because the agent’s core LLM model does not specialize on weather patterns, it obtains data from an external database including daily weather records for Greece over the last several years.

Despite receiving this additional information, the agent is still unable to predict the best weather conditions for surfing, thus the next subtask is established. For this subtask, the agent interacts with an external agent who specializes in surfing. Assume that while doing so, the agent discovers that high tides and bright weather with minimal to no rain give the optimum surfing conditions.

The agent may now use the information gathered from its tools to discover patterns. It can forecast which week next year in Greece will have high tides, sunny weather, and a low likelihood of rain. The results are then displayed to the user. The exchange of information amongst tools is what enables AI agents to be more versatile than standard AI models.

Learning and Reflection

AI agents increase their response accuracy by utilizing feedback mechanisms such as other AI agents and human-in-the-loop (HITL). Let’s go back to our prior surfing scenario to demonstrate this. After responding to the user, the agent retains the learnt knowledge as well as the user’s input in order to enhance performance and react to future user preferences.

If additional agents were used to achieve the aim, their input may also be utilized. Multi-agent feedback can be extremely effective in reducing the amount of time that human users spend providing instruction. However, users may offer feedback throughout the agent’s activities and internal reasoning to help align the results with the desired purpose.

Iterative refinement is the process of improving the logic and accuracy of an AI agent via feedback mechanisms. To prevent making the same mistakes, AI agents might keep data about prior hurdles in a knowledge base.

Agentic vs non-agentic AI chatbots

AI chatbots employ conversational AI methods like natural language processing (NLP) to interpret user inquiries and respond to them automatically. These chatbots are a modality, whereas agency is a technological foundation.

Non-agentic AI chatbots lack accessible tools, memory, and reasoning. They are limited to accomplishing short-term objectives and have no way forward. Non-agentic chatbots, which we know, require ongoing user input in order to reply. They can generate replies to typical prompts that are most likely consistent with user expectations, but they perform badly on inquiries specific to the person and their data. Because these chatbots have no recall, they are unable to learn from their errors if their replies are inadequate.

On the other hand, autonomous AI chatbots tend to adapt to user expectations over time, resulting in a more customized experience and thorough replies. They can execute difficult jobs without human involvement by breaking them down into smaller tasks and evaluating several options. These strategies may also be self-corrected and modified as required. Agentic AI chatbots, unlike non-agentic ones, evaluate their tools and utilize available resources to address knowledge gaps.

Reasoning paradigms

There isn’t a single standard architecture for developing AI bots. There are multiple models for addressing multistep issues.

ReAct: Reasoning and Action

The ReAct paradigm allows us to train agents to “think” and plan after each action and tool response in order to determine which tool to employ next. These Think-Act-Observe cycles are used to address problems step by step while iteratively improving replies.

Using the prompt structure, agents may be told to reason slowly and show each “thought”.4 The agent’s verbal thinking provides insight into how replies are constructed. In this paradigm, agents are constantly updating their context with fresh reasoning. This can be understood as a type of chain-of-thought prompting.

ReWOO (Reasoning WithOut Observation)

The ReWOO technique, unlike ReAct, eliminates the need for tool outputs for action planning. Instead, agents make their plans ahead of time. Redundant tool usage is prevented by predicting which tools to use when prompted by the user. This is ideal from a human-centered standpoint since the user may approve the plan before it is carried out.

The ReWOO process is comprised of three parts. In the planning module, the agent expects its future moves based on a user query. The next step is to collect the results provided by invoking these tools. Finally, the agent combines the basic plan with the tool outputs to create a response. This forward planning can significantly minimize token consumption, computational complexity, and the consequences of intermediary tool failure.

Different kinds of AI Agents

AI agents can be designed to have a variety of capacities. To save excessive computing complexity, a basic agent may be selected for easy goals. In order of simplest to most sophisticated, there are five primary agent types:

1. Simple reflex agents.

Basic reflexive agents are the most basic agent types that base their behaviors on perception. This agent has no memory and does not communicate with other agents when missing knowledge. These agents follow a set of so-called reflexes or rules. Thus, the tool is preprogrammed to do activities in response to specific criteria being satisfied.

If the agent faces a circumstance for which it is unprepared, it will be unable to respond effectively. The agents are only functional in fully visible contexts, allowing access to all relevant information.6

For example, a thermostat that activates the heating system at a predetermined time each night. The condition-action rule here states that if it is 8 p.m., the heating is engaged.

2. Model-based reflex agents.

Model-based reflex agents keep an internal model of the environment by combining their present perceptions and memories. As the agent receives new information, the model is modified. The agent’s actions are determined by its model, reflexes, prior precepts, and present condition.

In contrast to basic reflex agents, these agents may retain information in memory and function in partially visible and constantly changing environments. However, they are still constrained by their set of rules.

To exemplify, Consider a robotic vacuum cleaner. As it cleans a filthy room, it detects barriers such as furniture and adapts accordingly. To avoid being locked in a cycle of recurrent cleaning, the robot also retains a model of the regions it has previously cleaned.

3. Goal-based agents

Goal-based agents have both an internal representation of the environment and a goal or collection of goals. These agents look for action sequences that will help them achieve their goal and prepare these actions before carrying them out. This search and planning enhances their performance as compared to simple and model-based reflex agents.

For instance, Consider a navigation system that advises the shortest path to your location. The model evaluates multiple paths to attain your target, or aim. In this case, the condition-action rule indicates that if a faster route is discovered, the agent suggests it instead.

4. Utility-based agents

Utility-based agents choose the sequence of activities that will achieve the goal while simultaneously maximizing utility or reward. Utility is computed via a utility function. This function assigns a utility value, a metric that measures the usefulness of an action or how “happy” it will make the agent, to each situation based on a set of predefined parameters.

The criteria may include considerations such as progress toward the objective, time constraints, or computing complexity. The agent then chooses activities that maximize predicted utility. As a result, these agents are effective in situations where numerous scenarios meet the intended goal and the best one must be chosen.

For example, a navigation system that offers the best route to your location based on fuel economy, traffic time, and toll costs. This agent evaluates utility using this set of criteria to choose the best path.

5. Learning Agents

Learning agents have the same capabilities as other agent kinds but differ in their capacity to learn. New experiences are added to their basic knowledge base, which occurs independently. This learning improves the agent’s capacity to function in novel contexts. Learning agents can be utility or goal-based in their reasoning and are made up of four major components:

• Learning: This increases the agent’s understanding by allowing it to learn from its surroundings via precepts and sensors.
• Critic: This gives the agent feedback on whether the quality of their reply matches the performance requirement.
• Performance: This part is in charge of deciding what actions to take after learning.
• Problem generator: This generates numerous suggestions for activities to be performed.

For example, personalized recommendations on e-commerce websites. These agents save user action and preferences in their memory. This information is used to propose certain items and services to the user. The cycle is repeated every time fresh recommendations are presented. The user’s action is continually recorded for learning reasons. As a result, the agent’s accuracy gradually improves.

Use cases for AI agents

Customer experience.

AI agents may be integrated into websites and applications to improve the user experience by acting as virtual assistants, offering mental health assistance, mimicking interviews, and doing other similar duties.8 There are several no-code templates for user implementation, making the process of developing these AI agents even simpler.

Healthcare

AI agents have practical uses in healthcare. Multi-agent systems can be very beneficial for problem resolution in such scenarios. These technologies save medical workers time and effort on more critical activities, such as treatment planning for patients in the emergency department and medication process management.

Emergency Response

In the event of a natural disaster, AI agents can employ deep learning algorithms to extract the information of people on social networking platforms who require rescue. The locations of these users can be tracked to help rescue workers save more people in less time. As a result, AI agents can significantly improve human lives in both monotonous, repetitive jobs and life-saving emergencies.

Finance and Supply Chain

Agents may be programmed to assess real time financial data, predict future market trends, and improve supply chain management. The customization of autonomous AI agents allows us to receive outputs that are tailored to our specific data. When dealing with financial data, it is critical to implement security measures to protect data.

Benefits of AI Agents

Task automation

With continued improvements in generative AI and machine learning, there is an increasing interest in AI-powered workflow optimization, often known as intelligent automation. AI agents are AI technologies that can automate difficult operations that would normally require human intervention. This leads to achieving goals at a low cost, quick pace, and on a large scale. As a result of these developments, human agents no longer need to direct the AI assistant’s task creation and navigation.

Increased performance

Multi-agent frameworks often outperform single agents. This is because the more options an agent has for action, the more learning and reflection take place. An AI agent that incorporates knowledge and comments from other AI agents specialized in relevant fields might help with information synthesis. Agentic frameworks are unusual in that they allow AI agents to collaborate on the backend and fill information gaps, making them a strong tool and a significant leap in AI.

Quality of reaction

AI agents give more extensive, accurate, and tailored replies to users than typical AI models. This is highly important to us as consumers since higher-quality replies usually result in a better customer experience. As previously explained, this is accomplished by sharing information with other agents, utilizing external tools, and updating their memory stream. These habits occur spontaneously and are not planned.

Risks and Limitations

Multi-agent dependencies

Certain difficult tasks need the expertise of several AI agents. The orchestration of these multi-agent systems runs the risk of malfunctioning. Multi-agent systems based on the same foundation concepts may encounter common difficulties. Such flaws might result in a system-wide failure of all associated agents or expose vulnerabilities to malicious assaults. This emphasizes the necessity of data governance in the development of foundation models, as well as extensive training and testing procedures.

Infinite feedback loops

The ease of hands-off reasoning for humans employing AI agents is not without drawbacks. Agents that are unable to develop a thorough strategy or reflect on their results may find themselves constantly using the same tools, resulting in unending feedback loops. To avoid these redundancies, some form of real-time human monitoring may be employed.

Computational complexity

Developing AI agents from beginning is laborious and computationally costly. The resources necessary to train a highly effective agent can be significant. In addition, depending on the intricacy of the work, agents might need multiple days to finish it.

Data Privacy

If not properly managed, the integration of AI agents with corporate processes and customer management systems can generate major security risks. Assume AI bots are guiding the software development process, bringing coding copilots to the next level, or selecting pricing for clients with no human oversight or safeguards. The consequences of such a scenario might be negative due to agentic AI’s exploratory and frequently surprising behaviour. As a result, it is critical that AI providers like IBM, Microsoft, and OpenAI remain proactive and implement stringent security policies to guarantee sensitive employee and customer data is safely preserved.

Best Practices:

Activity Logs

To alleviate concerns about multi-agent dependency, developers can give users access to a log of agent operations. The steps may involve the usage of external tools and the description of external agents used to achieve the aim. This transparency gives customers insight into the iterative decision-making process, allows them to detect flaws, and fosters confidence.

Interruption

It is advised that autonomous AI bots do not run for extended periods of time. Particularly in the event of unanticipated endless feedback loops, shifts in availability to certain tools, or malfunctioning due to design errors. One approach to do this is to use interruptibility.

To maintain control, allow human users to gracefully interrupt a sequence of activities or the entire operation. Choosing whether and when to interrupt an AI agent needs considerable deliberation, since certain terminations might do more harm than benefit. For example, it may be better to allow a malfunctioning agent to continue aiding in a life-threatening situation rather than entirely shut it down.

Unique agent identifiers

Unique identities can be used to reduce the likelihood that agentic systems would be utilized maliciously. If these identities were necessary for agents to access external systems, it would be easier to track down the agent’s developers, deployers, and users. This would be especially useful if the agent used the information maliciously or caused unanticipated harm. This degree of accountability would make it safer for these artificial intelligence bots to operate.

Human oversight

To help AI bots learn in a new context, it can be beneficial to give some amount of human monitoring. This enables the AI agent to compare its performance to the expected level and modify accordingly. This type of input is useful for increasing the agent’s flexibility to user preferences.

Aside from that, it is excellent practice to get human consent before an AI agent conducts significant activities. For example, acts ranging from mass emailing to financial transactions should require human verification. Human monitoring is recommended in such high-risk settings.

---

Explore more blogs at: Jass Insights

---

Citation & Reference

Andrew Zhao, Daniel Huang, Quentin Xu, Matthieu Lin, Yong-Jin Liu, and Gao Huang, “Expel: Llm agents are experiential learners,” Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 38, No. 17, pp. 19632-19642, 2024, https://ojs.aaai.org/index.php/AAAI/article/view/29936

 Yonadov Shavit, Sandhini Agarwal, Miles Brundage, Steven Adler, Cullen O’Keefe, Rosie Campbell, Teddy Lee, Pamela Mishkin, Tyna Eloundou, Alan Hickey, Katarina Slama, Lama Ahmad, Paul McMillan, Alex Beutel, Alexandre Passos and David G. Robinson, “Practices for Governing Agentic AI Systems,” OpenAI, 2023, https://arxiv.org/pdf/2401.13138v3

 Tula Masterman, Sandi Besen, Mason Sawtell, Alex Chao, “The Landscape of Emerging AI AgentArchitectures for Reasoning, Planning, and Tool Calling: A Survey,” arXiv preprint, 2024, https://arxiv.org/abs/2404.11584

Gautier Dagan, Frank Keller, and Alex Lascarides, “Dynamic Planning with a LLM,” arXiv preprint, 2023. https://arxiv.org/abs/2308.06391

 Binfeng Xu, Zhiyuan Peng, Bowen Lei, Subhabrata Mukherjee, Yuchen Liu, and Dongkuan Xu, “ReWOO: Decoupling Reasoning from Observations for Efficient Augmented Language Models,” arXiv preprint, 2023, https://arxiv.org/abs/2305.18323

Sebastian Schmid, Daniel Schraudner, and Andreas Harth, “Performance comparison of simple reflex agents using stigmergy with model-based agents in self-organizing transportation.” IEEE International Conference on Autonomic Computing and Self-Organizing Systems Companion, pp. 93-98, 2021, https://ieeexplore.ieee.org/document/9599196

 Veselka Sasheva Petrova-Dimitrova, “Classifications of intelligence agents and their applications,” Fundamental Sciences and Applications, Vol. 28, No. 1, 2022.

Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, Wayne Xin Zhao, Zhewei Wei, and Jirong Wen, “A survey on large language model based autonomous agents,” Frontiers of Computer Science, Vol. 18, No. 6, 2024, https://link.springer.com/article/10.1007/s11704-024-40231-1

Jaya R. Haleema, Haleema, N. C. S. N. Narayana, “Enhancing a Traditional Health Care System of an Organization for Better Service with Agent Technology by Ensuring Confidentiality of Patients’ Medical Information,” Cybernetics and Information Technologies, Vol. 12, No. 3, pp.140-156, 2013, https://sciendo.com/article/10.2478/cait-2013-0031

Jingwei Huang, Wael Khallouli, Ghaith Rabadi, Mamadou Seck, “Intelligent Agent for Hurricane Emergency Identification and Text Information Extraction from Streaming Social Media Big Data,” International Journal of Critical Infrastructures, Vol. 19, No. 2, pp. 124-139, 2023, https://arxiv.org/abs/2106.07114

 Junyou Li, Qin Zhang, Yangbin Yu, Qiang Fu, and Deheng Ye. “More agents are all you need.” arXiv preprint, 2024, https://arxiv.org/abs/2402.05120

 Joon Sung Park, Joseph O’Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein, “Generative agents: Interactive simulacra of human behaviour,” Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, pp. 1-22, 2023, https://dl.acm.org/doi/10.1145/3586183.3606763

Alan Chan, Carson Ezell, Max Kaufmann, Kevin Wei, Lewis Hammond, Herbie Bradley, Emma Bluemke, Nitarshan Rajkumar, David Krueger, Noam Kolt, Lennart Heim and Markus Anderljung, “Visibility into AI Agents,” The 2024 ACM Conference on Fairness, Accountability, and Transparency, pp. 958-973, 2024, https://arxiv.org/abs/2401.13138

 Devjeet Roy, Xuchao Zhang, Rashi Bhave, Chetan Bansal, Pedro Las-Casas, Rodrigo Fonseca, and Saravan Rajmohan, “Exploring LLM-based Agents for Root Cause Analysis,” arXiv preprint, 2024, https://arxiv.org/abs/2403.04123

---