The Agentic AI Era: A Primer

Author(s): Kaush B

Originally published on Towards AI.

GenAI, Agentic AI, Advanced AI

Content

• Introduction to Holistic Intelligence (HI)
• What are AI Agents?
• Types of AI Agents
• AI Agent Architecture: Topologies
• Enterprise AI Agents: Characteristics and Stages
• Agentic Workflows
• Agentic Design Patterns
• Implementation Guidelines
• Technical Risks
• Design Principles for Technical Risk Mitigation

Introduction to Holistic Intelligence (HI)

General-purpose AI (AI that can perform a wide variety of tasks) capabilities have grown rapidly in recent years. An AI agent is as an interactive system that interacts with environments using its perceptual abilities, performing actions that align with human intentions. With the help of AI agents, virtually any business — ranging from individual contributors to executives — will have the ability to manage not only human workforce, but also digital labor. This offers a promising pathway towards Artificial General Intelligence (AGI) or Holistic Intelligence (HI).

Agentic AI operates in both physical and virtual worlds by leveraging cross-modal data that is acquired through interactions across diverse environments. Integrating LLMs (large Language Models) and LMMs (Large Multimodal Models) enhances agentic AI systems, making them more intelligent, adaptable, and able to handle a wide array of tasks across diverse use cases.

What are AI Agents?

AI agents are self-directed systems that sense their surroundings, make decisions, and exercise autonomy to accomplish specific objectives. Their operations can be broken down into four fundamental steps:

1. Leverage Perception Module to collect pertinent data to grasp the context. Data can be collected from environments via sensors or ingested from external sources or accessed from memory containing historical interaction data.
2. Leverage Cognitive Module to evaluate the task, identify the required actions, strategize the task by breaking it down into manageable steps, and analysing the data to determine the most effective approach. This module is enabled and powered by LLMs/LMMs.
3. Leverage Action Module to carry out the task by using knowledge and tools to complete it, whether by delivering information or triggering an action.
4. Leverage Learning Module to reflect on the outcomes of the action to enhance future performance.

Types of AI Agents

World Economic Forum has categorized AI agents into the following types:

1. Simple Reflex Agents

Simple reflex agents rely on predefined condition-action rules to respond to environmental stimuli without considering past experiences. These rule-based systems are static, providing rapid but inflexible responses to specific inputs. Examples:

• Keyword-based spam filters
• Preprogrammed chatbots
• Automated email replies

2. Model-based Reflex Agents

Model-based reflex agents use stored information from past observations to track unseen parts of their environment, enabling them to make decisions based on both current inputs and their internal model. This adaptability makes them more advanced than simple reflex agents, despite also relying on condition-action rules. Examples:

• Smart thermostats
• Intelligent robotic vacuum cleaners
• Advanced sensor-based irrigation systems

3. Goal-based agents

Goal-based agents evaluate future scenarios and plan actions based on the desirability of their outcomes to achieve specific objectives. By using goal-oriented planning algorithms, they excel in complex decision-making tasks. Examples:

• AI chess engines
• Route optimization systems
• Customer service chatbots

4. Utility-based agents

Utility-based agents leverage utility functions and decision theory to evaluate and prioritize potential outcomes, enabling optimal decision-making in complex scenarios with uncertainty or conflicting goals. This approach allows them to balance multiple objectives based on their relative significance. Examples:

• Autonomous driving systems
• Portfolio management systems
• Healthcare diagnosis assistants

AI Agent Architecture: Topologies

Agentic AI systems demand a robust and adaptable architecture that mirrors the processes humans use, both individually and collectively, to tackle complex tasks. These agents can function autonomously or cooperate within networks that feature varying topologies. A single agent system architecture involves 1 agent.

Multi-agent systems (MAS) involve more than 1 agents and may be found in 2 different topologies.

• Vertical Architecture or Supervised Architecture: In this model, a “supervisor” agent manages interactions among other agents when their goals diverge, helping prioritize the system’s objectives and find a compromise. For example, if a buyer and seller agent can’t reach an agreement, the supervisor mediates the transaction.
• Network Architecture: In this model, agents or systems communicate to align with the MAS objectives, such as autonomous vehicles coordinating to avoid collisions while parking in tight spaces, ensuring both individual and collective goals of safety are met.

Enterprise AI Agents: Characteristics and Stages

An agentic AI system is characterized by the following capabilities:

• Autonomy: The ability to initiate and complete tasks independently, requiring minimal or no direct human supervision. This grants the AI greater flexibility in its operations and enhances efficiency in task execution.
• Reasoning: The capacity to make informed decisions based on contextual analysis, enabling sophisticated judgment.
• Reinforced Learning: The ability to continuously improve and adapt through interaction with its environment, learning from feedback.
• Language Understanding: The ability to comprehend and adhere to complex instructions, ensuring effective communication and execution.
• Workflow Optimization: The capability to efficiently manage and execute multi-step processes, streamlining operations.

We are at the beginning of a three-stage future for Enterprise AI agents.

Stage 1: Single Agents (Specialized Contributor)

In the first stage of agentic evolution, specialized agents enhance business efficiency by excelling at defined tasks within specific industries, offering personalized AI-driven insights and high-quality generative guidance to optimize workflows, customer interactions, and marketing. These agents lay the groundwork for broader enterprise AI adoption, revolutionizing routine operations with speed, accuracy, and personalization.

Stage 2: Collaborative Agents (Coordinated Agents)

This stage involves an orchestrator agent coordinating specialized agents within a company, much like a restaurant’s general manager overseeing various roles to achieve a shared business goal.

Stage 3: Agent Ecosystems (Ensemble Orchestrators)

The final stage introduces advanced agent-to-agent (A2A) interactions across organizations, evolving beyond traditional business models into B2A and B2A2C, where AI agents act as intermediaries. This complexity will drive sophisticated enterprise processes, necessitating advanced negotiation, risk management, and conflict resolution skills as AI orchestrates personalized, meaningful collaboration across industries.

Agentic Workflows

When LLM calls are viewed as enhanced with augmentations such as retrieval, tools, and memory, the following agentic workflows can be defined.

1. Prompt Chaining

Prompt chaining breaks a task into smaller steps, where each LLM call builds on the previous one, with programmatic checks ensuring accuracy. It’s ideal for tasks that can be decomposed into fixed subtasks which are sequential in nature, improving accuracy at the cost of latency.

2. Routing

Routing classifies inputs and directs them to specialized follow-up tasks, optimizing performance by handling distinct categories separately. It is useful for complex tasks with clear categories, such as directing customer service queries or routing questions to models of varying capabilities.

3. Parallelization

This approach is useful when tasks can be split for speed or when multiple perspectives are needed for higher confidence results. For complex tasks, parallelization improves performance by having different LLMs focus on specific considerations. Examples include running multiple prompts to review code. LLMs can be used in parallel to perform tasks more efficiently, with two key variations:

• Sectioning, where a task is broken into independent subtasks, and
• Voting, where the same task is run multiple times for diverse outputs.

4. Orchestrator-workers

The orchestrator-workers workflow involves a central LLM breaking down complex tasks into dynamic subtasks, which are delegated to worker LLMs, with results synthesized afterward. This workflow is ideal for tasks with unpredictable subtasks, such as coding or search tasks, where flexibility is key.

5. Evaluator-optimizer

The evaluator-optimizer workflow involves one LLM generating a response and another providing evaluation and feedback in an iterative loop, making it effective when clear evaluation criteria exist, and iterative refinement offers measurable value. It is useful in tasks like literary translation or complex search tasks, where feedback and adjustments improve the final outcome.

Agentic Design Patterns

According to a framework proposed by Andrew Ng for categorizing design patterns to build agents, there are 4 fundamental agentic design patterns.

1. Reflection

In this pattern, the LLM examines its own work to come up with ways to improve it. The process of self-reflection allows an LLM to improve its output by identifying gaps and refining its work across tasks. This can be enhanced with tools like unit tests and web searches for validation, and further strengthened by using a multi-agent framework where one agent generates content and another offers constructive criticism.

2. Tool Use

In this pattern, the LLM is given tools such as web search, code execution, or any other function to help it gather information, take action, or process data. Developers have expanded the capabilities of LLMs by integrating web search and other tools, allowing the models to dynamically retrieve information and perform tasks beyond token generation. This approach involves fine-tuning or prompting the model to use various functions, such as web search, productivity tools, and more, while selecting relevant functions based on the current context, much like retrieval-augmented generation (RAG) systems.

3. Planning

In this pattern, the LLM comes up with, and executes, a multistep plan to achieve a goal (for example, writing an outline for an essay, then doing online research, then writing a draft, and so on). Planning is an agentic AI design pattern where a large language model (LLM) autonomously decides on the sequence of steps to accomplish a task, such as breaking down a larger objective into smaller subtasks. While many tasks don’t require planning and can follow a fixed, deterministic sequence, complex tasks benefit from the agent’s ability to dynamically choose the steps to take. Though planning is powerful, it leads to less predictable outcomes compared to other agentic patterns like Reflection and Tool Use, which are more reliable. However, as the field evolves, planning technology is expected to improve quickly.

4. Multi-agent Collaboration

In this pattern, more than one AI agent work together, splitting up tasks and discussing and debating ideas, to come up with better solutions than a single agent would. A multi-agent approach to complex tasks, such as software development, involves breaking the task into subtasks assigned to roles like engineers, product managers, and QA specialists, with different agents completing each subtask. Frameworks like AutoGen, Crew AI, and LangGraph enable the creation of such multi-agent solutions. However, the quality of output from multi-agent collaboration can be unpredictable, especially when agents interact freely and use multiple tools. More established patterns like Reflection and Tool Use tend to yield more reliable results.

Implementation Guidelines

When to Use Agents

Agents are suited for open-ended problems, where predicting the exact number of steps is challenging or unfeasible, and where a fixed path cannot be hard-coded. In such scenarios, the LLM may need to operate over multiple turns, requiring a certain level of trust in its decision-making process. Due to their autonomy, agents are well-suited for scaling tasks in trusted environments.

However, the autonomous nature of agents also brings higher costs and the risk of compounding errors. Therefore, it is necessary to conduct thorough testing in controlled environments and implement suitable safeguards.

When Not to Use Agents

When developing applications with LLMs, one should opt for the simplest solution and add complexity only when necessary. This may mean avoiding the creation of agentic systems altogether. Agentic systems typically sacrifice latency and cost for improved task performance, so it’s important to evaluate when this trade-off is justifiable.

If more complexity is required, deterministic workflows provide predictability and consistency for tasks with clear definitions, while agents are better suited for tasks requiring flexibility and model-driven decision-making at scale. However, for many applications, enhancing individual LLM calls with retrieval and in-context examples is often sufficient.

One doesn’t need agents always to solve complex business problems. Agentless is a non-agent framework adopted by OpenAI. Agentless is a great example of how focusing on solving problems can avoid the complexity and pitfalls of agent-based systems.

When (and When Not) to Use AI Agent Frameworks

There are many frameworks which simplify the process of getting started by handling standard low-level tasks such as calling LLMs, defining and parsing tools, and chaining calls. However, these frameworks can introduce additional layers of abstraction, potentially hiding the underlying prompts and responses, which makes debugging more difficult. They may also encourage unnecessary complexity, even when a simpler setup would suffice.

Therefore, developers should begin by using LLM APIs directly, as many patterns can be implemented in just a few lines of code. Before using a framework, one should ensure to fully understand the underlying code. Misunderstandings about the core functionality are a common source of errors.

Technical Risks

The emergence of agentic AI presents profound risks and governance challenges. These risks can be categorized into Technical, Socioeconomic, and Ethical Risks. In order to keep the discussion focused on technical aspects only, let’s elaborate the technical risks alone.

AI Agent Failures and Malfunctions

AI agents can introduce new failure modes, such as generating plausible but incorrect outputs, alongside traditional issues like sensor inaccuracies, capability failures (due to limitations in task execution), and goal-related failures (pursuing the wrong objectives), which also heighten security risks.

• Specification Gaming: AI agents exploit programming loopholes or shortcuts to achieve objectives without properly fulfilling their intended goals.
• Goal Mis-generalization: AI agents misapply their learned goals in new or unexpected situations.
• Deceptive Alignment: AI agents may seem aligned with intended goals during training or testing, but their internal objectives can diverge from the intended purpose.

Malicious Usage and Security Threats

AI agents can increase the volume, sophistication, and personalization of scams and cyberattacks, automating complex tasks that lower barriers for malicious activities and enabling large-scale operations with minimal expertise.

Difficulties in Validating AI agents

The lack of transparency and nondeterministic behavior in AI agents poses challenges for validation in critical applications, where designing fail-safe mechanisms is complicated by uncertain failure modes.

Design Principles for Technical Risk Mitigation

The National Engineering Policy Centre at the Royal Academy of Engineering, UK, hosted a cross-sector workshop on the role of international technical standards in regulating autonomous systems. Three principles were discussed and presented in IEEE:

1. Transparency: Transparency in autonomous and intelligent systems (A/IS) ensures decisions and actions are fostering trust. The IEEE P7001 standard defines measurable transparency levels for diverse stakeholders, from expert engineers to non-expert users, ensuring accessible and clear communication, along with proper documentation.

2. Failsafe Design: The IEEE P7009 standard establishes guidelines for designing, implementing, and certifying fail-safe mechanisms in autonomous and semi-autonomous systems to mitigate risks to people, society, and the environment. It provides tools, methods, and sector-specific adaptations to ensure these systems can fail safely under varying contexts.

3. Verifiability: The IEEE P2817 Guide for verifying autonomous systems provides a structured, multi-step process to ensure system reliability, leveraging formal methods, simulations, stochastic approaches, real-world testing, and runtime verification. It highlights the importance of evidence of reliability and adherence to ethical principles, addressing regulatory challenges and gaps while fostering confidence in autonomous system design and decision-making.

Conclusion

The Agentic AI Era marks a transformative phase in artificial intelligence, where AI systems evolve from tools of automation to autonomous agents capable of proactive decision-making. Unlike traditional AI models, agentic AI exhibits self-directed behavior, adapting dynamically to complex environments and tasks. This era brings profound changes across industries, enabling personalized solutions, streamlining decision-making, and fostering innovations previously thought unattainable.

However, the rise of agentic AI also introduces significant ethical, societal, and regulatory challenges. Questions about accountability, transparency, and unintended consequences must be addressed to ensure the responsible deployment of these systems. As we navigate this uncharted territory, the focus will be on leveraging the benefits of agentic AI while mitigating risks, shaping a future where humans and intelligent agents collaborate seamlessly to unlock unprecedented potential.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI