Open in app Sign up Sign in

Search

5 Levels Of AI Agents (Updated)

𝗔𝘂𝘁𝗼𝗻𝗼𝗺𝗼𝘂𝘀 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁𝘀 are AI systems capable of performing a series of complex
tasks independently to achieve a goal.

Cobus Greyling · Follow

7 min read · Oct 11, 2024
 Listen Share

The phrases AI Agents, Autonomous Agents, Agentic Application, or what I refer to as

Agentic X are all terms which are used interchangeably.

Some Background
I love the example from Agile’s rise in organisations, where project managers
evolved into Scrum Masters, adapting to iterative development cycles.

Similarly, Conversational AI has transformed, shifting from basic chatbot

frameworks to advanced prompt-engineering tools, and now into comprehensive
AI Agent builders.

Understanding the components of an agentic framework is crucial in leveraging

these advancements effectively.

In this article, I use the terms RPA, chaining, prompt chaining, and chatbot dialog flows
interchangeably, as they all refer to a similar approach of using predefined, sequential
nodes to guide processes.

What are Agents? In Short…

Agentic applications leverage one or more Language Model as their core
foundation or backbone, dynamically generating responses and actions.

These applications manage states and transitions while constructing event

chains in real-time to address specific user queries, providing adaptive
solutions.

Agents excel at handling ambiguous or implicit questions, breaking them into

sequential sub-steps and iteratively processing through cycles of action,
observation, and reflection until reaching a final resolution.

Latency and cost management are critical for conversational implementations,

balancing responsiveness with resource efficiency. Latency can be problematic
with Agentic implementations.

Inspectability and observability are essential for production implementations,

with robust mechanisms developed to reveal the states and pathways traversed
by agents, ensuring transparency.
 To accomplish tasks, agents have access to diverse tools, each equipped with a
clear purpose — whether it’s making API calls, conducting calculations, or
searching the web.

Human-in-the-Loop (HITL) can serve as an auxiliary tool, enabling agents to

seek human input when needed, expanding their operational capabilities.

New agent tools can be seamlessly integrated to extend capabilities, allowing for
continuous adaptation and enhancement of autonomous agent functions.

Agents possess true autonomy, making decisions and performing actions

independently, requiring minimal human supervision. Levels of autonomy are
set by the number of iterations an AI Agent can cycle through, in order to reach
a conclusion; and the number of tools at its disposal.

With advanced flexibility, agents dynamically select and sequence tools based
on situational needs, employing reasoning and adaptive strategies to solve
complex tasks as they arise.

22 Key Differences Between AI Agents & Traditional Chaining/RPA

Below is an in-depth comparison of how AI Agents stand apart from traditional
chaining and Robotic Process Automation (RPA) methods based on the criteria
shown…

1. Flexibility, Autonomy, Reasoning

AI Agents: Exhibit high flexibility and autonomy, capable of complex reasoning and
decision-making based on context. They can adapt to unforeseen situations,
respond to changing data, and make real-time adjustments.

Chaining/RPA: Typically follow predefined rules and sequences with limited

flexibility. They execute tasks as they were programmed without deviation, making
them less adaptable to changes in their environment.

2. Granular State-Based
AI Agents: Operate with a granular, state-based approach, maintaining an internal
and dynamic state and understanding of the environment. This allows them to track
changes over time and adjust their actions accordingly.

Chaining/RPA: Generally lack granular state awareness and operate on fixed

workflows. They are less adept at handling complex or evolving tasks that require
contextual awareness.

3. RPA Approach
AI Agents: Use machine learning and natural language processing to make
decisions dynamically, going beyond rule-based automation.

Chaining/RPA: Rely heavily on traditional automation techniques, such as screen

scraping and hardcoded rules, which are inflexible and require significant
reconfiguration for any changes.

4.Human-in-the-Loop (HITL)
AI Agents: Often incorporate HITL for complex tasks or when encountering
uncertain situations, allowing human intervention to guide the process or provide
feedback.

Chaining/RPA: May involve humans for exception handling, but usually this is not a
built-in feature. HITL is less integrated compared to AI agents. Tradition
chatbots/chaining usually follow method of a complete transfer to an agent if the
chatbot does not fulfil the intent.

5. Managing Cost
AI Agents: Can be costly to deploy and maintain due to resource requirements,
though their adaptability and efficiency can lead to long-term cost savings.
Chaining/RPA: Often lower upfront costs, especially for simple repetitive tasks, but
can become expensive if frequent updates and maintenance are needed.

6. Optimising Latency
AI Agents: Utilise optimisation strategies to minimise latency, often by prefetching
data, parallel processing, or making real-time adjustments. But latency is often hard
to improve for reasoning and decomposition tasks.
Chaining/RPA: May experience latency due to rigid workflows and sequential
processing, with limited real-time optimisation. However, in general easier to
optimise.
7. LLM-Generated Action Sequence
AI Agents: Use Language Models to generate action sequences dynamically,
enabling them to handle complex, multi-step tasks based on evolving context.
Chaining/RPA: Action sequences are predefined and do not benefit from LLM-
driven flexibility, limiting their ability to handle nuanced or conversational tasks.

8. Seamless Tool Integration

AI Agents: Often integrate various tools and services seamlessly, including APIs,
databases, and external resources, to enhance functionality dynamically.
Chaining/RPA: Integration is typically more rigid, requiring manual configuration
and less dynamic adaptation to new tools or services.

9. Explainability / Observability / Inspectability

AI Agents: Often include features for explainability and observability, providing
insights into decision-making processes, which is critical for trust and compliance.
But in most cases this aspect lacks.
Chaining/RPA: The sequence of events are set for different workflows.
10. Design Canvas Approach
AI Agents: Often rely on traditional programming environments with less focus on
visual workflow design, making complex task configuration more challenging.
Chaining/RPA: May use design canvases for configuring complex workflows, often
visually represented, allowing for intuitive adjustments and reconfiguration.

11. Conversational Oriented

AI Agents: Can engage in conversational tasks, using natural language
understanding to interact with users effectively.
Chaining/RPA: Generally designed for conversational interfaces (considering
traditional chatbot frameworks), though they can interact with simple text-based
inputs.
12. Adaptive Learning Capabilities
AI Agents: can learn from new data and experiences over time, allowing them to
improve autonomously, unlike traditional RPA which relies on pre-set rules and
lacks learning capabilities.

13. Contextual Awareness

AI Agents have the ability to understand and adapt to the context of interactions,
which enhances decision-making and response accuracy. RPA, chatbot flows, on the
other hand, operates in a static, predefined context.

14. Dynamic Task Decomposition

AI Agents can break down complex tasks into smaller, manageable subtasks
dynamically and adjust these based on real-time feedback. RPA/Chatbots/Prompt
Chaining typically follows a linear, fixed sequence of tasks.

15. Real-Time Decision Making

AI Agents can make decisions on-the-fly based on live data, whereas RPA follows a
preset decision-making process based on pre-programmed logic.
16. Unstructured Data Handling
AI Agents are capable of understanding and processing unstructured data, like
natural language and images, through AI models, while RPA and workflows usually
work with structured data which is well defined and classified.

17. Goal-Oriented Behaviour

AI Agents work towards high-level objectives and can modify their approach as
needed, whereas Chaining scripts are designed to achieve specific tasks without
overarching goals.
18. Scalability in Diverse Environments
AI Agents can be deployed in a wide range of environments and scale easily with
minimal configuration changes, while RPA/workflows solutions may require
significant customisation to adapt to different platforms or systems.

19. Proactive Engagement

AI Agents can initiate actions and engage proactively based on user behaviours or
external triggers. RPA/workflows/chains are more reactive, executing tasks only
when prompted by a specific event.

20. Tool Interoperability and API Flexibility

AI Agents are often designed to work seamlessly with a variety of tools and APIs,
adapting as needed, whereas RPA/chains/workflows/chatbot solutions are generally
more rigid and specific to certain tools or systems.

21. No Low-Code IDEs

AI Agents: development environments are typically more pro-code at this stage.
Where RPA/Workflows/chaining/Chatbots are more established in their no-code
design canvas approach to building.
22. Dynamic Adaptability to Unseen Scenarios
AI Agents: Can adapt to new and unseen scenarios by leveraging machine learning
and contextual understanding, which allows them to make decisions even in cases
they have not explicitly been trained for. This makes them highly adaptable to
changing environments or unexpected user inputs.

Chaining/RPA: Typically struggle with scenarios outside their predefined scripts or

rules. They operate based on specific sequences and can fail or require manual
intervention when faced with unexpected situations or edge cases they were not
designed to handle.

Follow me on LinkedIn ✨✨

Chief Evangelist @ Kore.ai | I’m passionate about exploring the intersection of AI and
language. From Language Models, AI Agents to Agentic Applications, Development
Frameworks & Data-Centric Productivity Tools, I share insights and ideas on how these
technologies are shaping the future.

COBUS GREYLING
Where AI Meets Language | Language Models, AI Agents, Agentic
Applications, Development Frameworks & Data-Centric…
www.cobusgreyling.com

Get an email whenever Cobus Greyling publishes.

Get an email whenever Cobus Greyling publishes. By signing up, you
will create a Medium account if you don't already…
cobusgreyling.medium.com

AI Artificial Intelligence Ai Agents Agentic Applications Machine Learning

Follow

Written by Cobus Greyling

26K Followers · 0 Following

I’m passionate about exploring the intersection of AI & language. www.cobusgreyling.com

Responses (10)
 What are your thoughts?

Respond

Karthikeyan V
Oct 13, 2024

Point 2 granular state appears to contradict between the image and the text. please clarify / correc it.

49 Reply

Oliver Parker
Oct 17, 2024 (edited)

nice overview. also can explore interactive ai agents market landscape map grouped by categories

https://aiagentsdirectory.com/landscape

19 Reply

Alex Amaya
Oct 14, 2024

Nice comparison between RPAs and AI agents, worth incorporating some of these concepts on our sales
pitch, thank you

53 Reply

See all responses

More from Cobus Greyling

 Cobus Greyling

How Did We Go From Chatbots To AI Agents?

The chatbot ecosystem was severely impacted by the advent of Language Models and
Generative AI. Incumbents had to figure out how to adapt…

Jan 6 90

Cobus Greyling

When to Use Functions, a Multi-Tool AI Agent, or Multiple Agents

Sometimes, a single agent with multiple tools is enough — or even just functions.
6d ago 100 1

Cobus Greyling

Transforming Industries with Vertical AI Agents

A recent study explored the pivotal role AI Agents will play in various industry verticals. It also
introduced the concept of Cognitive…

Jan 8 58 3

Cobus Greyling

An AI Agent Architecture & Framework Is Emerging

We are beginning to see the convergence on fundamental architectural principles that are
poised to define the next generation of AI agents…

Sep 16, 2024 608 7

See all from Cobus Greyling

Recommended from Medium

In Towards AI by Christopher Tao

Do Not Use LLM or Generative AI For These Use Cases

Choose correct AI techniques for the right use case families

Aug 10, 2024 4.1K 42

 In Towards Data Science by Eric Broda

Agentic Mesh: The Future of Generative AI-Enabled Autonomous Agent

Ecosystems
Agentic Mesh is an ecosystem that lets Autonomous Agents find each other, collaborate,
interact, and transact in a safe and trusted manner.

Nov 6, 2024 1.3K 25

Lists

Natural Language Processing

1909 stories · 1568 saves

Predictive Modeling w/ Python

20 stories · 1811 saves

AI Regulation
6 stories · 679 saves

Generative AI Recommended Reading

52 stories · 1631 saves
 In Vedcraft by Ankur Kumar

Building Intelligent Apps with Agentic AI: Top Frameworks to Watch for in
2025

Jan 24 121 4

In AI Advances by Kenny Vaneetvelde

Want to Build AI Agents? Tired of LangChain, CrewAI, AutoGen & Other

AI Frameworks? Read this!
Frameworks like LangChain, CrewAI, and AutoGen have gained popularity by promising high-
level abstractions for building AI systems. Yet…
 Jan 19 1.1K 24

Alberto Romero

DeepSeek Is Chinese But Its AI Models Are From Another Planet

OpenAI and the US are in deep trouble

Jan 22 5.6K 177

In Level Up Coding by Ida Silfverskiöld

Agentic AI: Build a Tech Research Agent

Using a custom data pipeline with millions of texts

Sep 5, 2024 1K 16

See more recommendations