🤖What are Agentic Workflows?

If you’ve ever tried to write a long paper, you know it’s not a linear process. You don’t just sit down and type it out from start to finish. Instead, you write an outline, then a section, then another, and eventually, you go back and revise. It’s iterative. You refine as you go. This is how humans naturally approach complex tasks. So why don’t we design AI systems to work the same way?

This is the idea behind agentic workflows, a concept I came across while reading blog posts by Andrew Ng, a well-known AI researcher and educator. Ng, who co-founded Google Brain and DeepLearning.AI, has been exploring how large language models (LLMs) can be made more effective by mimicking human problem-solving processes. He outlines four key workflow patterns that improve LLM performance: reflection, tool use, planning, and multi-agent collaboration. Let’s break these down.

Reflection: The Self-Improving Loop

The first pattern is reflection. Here, the LLM critiques and revises its own work. For example, if you ask it to write a paper, it might start by generating an outline, then a draft, and finally, a polished version. But instead of stopping there, it can review its own output, identify weaknesses — like factual errors or awkward phrasing — and refine the text. This self-reflective loop often leads to better results than a single pass. It’s like having an editor inside the model, constantly pushing it to improve.

Tool Use: Expanding Capabilities

The second pattern is tool use. LLMs are powerful, but they’re not omniscient. By giving them access to external tools — like web search, calendars, code repositories, or APIs — they can gather new information, process data, and interact with other systems. This makes them more versatile. Instead of relying solely on their internal knowledge, they can pull in real-time data or execute tasks that require external resources. Think of it as equipping the model with a Swiss Army knife.

Planning: Breaking Down Complexity

The third pattern is planning. Humans don’t tackle complex tasks all at once. We break them into smaller steps. The same goes for LLMs. In this workflow, the model first generates a multi-step plan to achieve its goal. For instance, if the task is writing a paper, the plan might include creating an outline, drafting sections, and revising. The model then executes these steps one by one, adjusting as it goes based on feedback. This structured approach helps it handle tasks that would otherwise be overwhelming.

Multi-Agent Collaboration: Teamwork for AI

The fourth pattern is multi-agent collaboration. Instead of relying on a single LLM, multiple agents work together, either by debating ideas or dividing tasks. This collaborative approach leverages the strengths of different agents, leading to more nuanced and comprehensive solutions. It’s like having a team of experts, each contributing their unique perspective.

Why This Matters

These agentic workflows make LLMs more adaptive and autonomous. They allow models to approach multi-step tasks with the same depth and efficiency as humans. This is a big deal. Most AI systems today are designed to perform tasks in a single pass. They generate an answer, and that’s it. But complex problems often require iteration, refinement, and collaboration. By mimicking how humans work, agentic workflows enable LLMs to tackle these problems more effectively.

What’s Next?

Ng’s ideas are already influencing how we think about AI systems. For example, imagine an LLM that not only writes code but also debugs it, tests it, and refines it based on feedback. Or a model that can plan a project, delegate tasks to other agents, and adjust the plan as new information comes in. These are the kinds of systems agentic workflows make possible.

But there’s still a lot to explore. How do we design these workflows to be efficient and scalable? How do we ensure they’re reliable and safe? These are open questions, and they’re worth thinking about if you’re interested in the future of AI.

Final Thoughts

One of the most striking things about agentic workflows is how obvious they seem in hindsight. Of course AI systems should work like humans. Of course they should iterate, use tools, plan, and collaborate. But it took someone like Andrew Ng to articulate this clearly and show how it can be done.

If you’re working with AI, these ideas are worth paying attention to. They’re not just theoretical — they’re practical. And they’re likely to shape how we build and use AI systems in the years to come.

So the next time you’re writing a paper or tackling a complex project, think about how you approach the task. Chances are, you’re using something like an agentic workflow. And if you are, maybe your AI systems should too.

Further Reading:

🤖ChatGPT for Vulnerability Detection by Tahir Balarabe

*🚀DeepSeek R1 Explained: Chain of Thought, Reinforcement Learning, and Model Distillation*

What are AI Agents?

⚙️LangChain vs. LangGraph: A Comparative Analysis

Stable Diffusion Deepfakes: Creation and Detection

🔗What is Model Context Protocol? (MCP) Architecture Overview

The Difference Between AI Assistants and AI Agents (And Why It Matters)

⚛️Microsoft’s Majorana 1: A Scalable Quantum Computing Breakthrough

🤖DeepSeek R1 API Interaction with Python

Frequently Asked Questions on Agentic Workflows for Large Language Models (LLMs)

1. What are agentic workflows in the context of Large Language Models (LLMs)?

Agentic workflows refer to methods that enable LLMs to operate more autonomously and iteratively, similar to how humans tackle complex tasks. Instead of a single pass generation, these workflows incorporate strategies that allow LLMs to refine their outputs through self-critique, utilization of external tools, structured planning, and collaboration with other agents. The goal is to enhance the overall performance of LLMs in tasks like coding, text generation, and decision-making by mimicking human problem-solving processes.

2. What are the four key LLM workflow patterns that constitute agentic workflows, as highlighted by Andrew Ng?

Andrew Ng identified four key patterns that contribute to improved LLM performance within agentic workflows:

• Reflection: The LLM critiques and revises its own work to identify and address weaknesses, such as factual inaccuracies or structural issues, leading to improved outputs through iterative refinement.
• Tool Use: The LLM is equipped with the ability to access and utilize external tools like web search engines, calendars, code repositories, code execution environments, or APIs. This allows the LLM to gather new information, process data, and interact with external systems to enhance its problem-solving capabilities beyond its internal knowledge.
• Planning: The LLM first generates a multi-step plan to achieve a complex goal, breaking it down into smaller, manageable steps. It can then execute these steps sequentially and adjust the plan as needed based on feedback or intermediate results.
• Multi-Agent Collaboration: Multiple LLM agents work together to accomplish a task. This can involve debating different ideas, splitting the workload based on individual agent capabilities, and integrating their outputs to achieve a more comprehensive and nuanced solution.

3. How does the concept of human problem-solving inspire agentic workflows for LLMs?

Agentic workflows are inspired by the way humans approach complex projects. For instance, when writing a lengthy paper, individuals typically don’t write it linearly from start to finish in one go. Instead, they engage in an iterative process involving outlining, drafting sections, reviewing, rewriting, and reorganizing. This involves self-reflection, potentially using external resources (like research papers), planning the structure, and even collaborating with others for feedback. Agentic workflows aim to replicate this iterative and multi-faceted approach in LLM-driven systems, allowing them to refine their work and leverage various capabilities to solve complex, multi-step tasks more effectively.

4. Can you elaborate on the “Reflection” workflow pattern and its benefits?

The reflection workflow pattern involves prompting the LLM to not only generate an output (like a piece of text or code) but also to subsequently review and critique its own work. This self-assessment allows the LLM to identify shortcomings such as factual errors, structural weaknesses, or awkward phrasing. Based on this self-critique, the LLM can then generate an improved version of its output, addressing the identified issues. This cycle of generation and self-reflection can be repeated multiple times, leading to a progressively refined and higher-quality final result. The benefit of reflection is that it allows the LLM to autonomously identify and correct its own mistakes, leading to better outcomes than a single, unreviewed generation.

5. How does the “Tool Use” workflow pattern enhance the capabilities of LLMs?

The “Tool Use” workflow pattern significantly expands the capabilities of LLMs by allowing them to interact with the external world. By integrating access to tools like web search, code execution environments, or specific APIs, LLMs are no longer limited to their internal training data. They can actively gather up-to-date information, perform computations, interact with external systems, and validate their outputs. For example, an LLM tasked with answering a question requiring current information can use a web search tool to retrieve relevant data before formulating its response. This ability to leverage external resources makes LLMs more versatile and capable of tackling a wider range of tasks that require real-time data or specific functionalities.

6. What does the “Planning” workflow pattern entail and why is it important for complex tasks?

The “Planning” workflow pattern involves the LLM first developing a structured, multi-step plan to achieve a given objective. For complex tasks that cannot be completed in a single step, the LLM breaks down the overall goal into smaller, more manageable sub-tasks. Using the example of writing a paper, the planning stage might involve creating an outline, deciding on the content of each section, and determining the order of writing and review. Once the plan is generated, the LLM can then execute each step sequentially, potentially adjusting the plan based on feedback or intermediate results. This structured approach is crucial for complex tasks as it provides a roadmap for the LLM, allowing it to approach the problem in a more organized and efficient manner, increasing the likelihood of a successful outcome.

7. How does “Multi-Agent Collaboration” contribute to improved LLM performance?

The “Multi-Agent Collaboration” workflow pattern leverages the idea that different agents (which could be different instances of an LLM or specialized LLMs) can bring different strengths and perspectives to a problem. By having multiple agents work together, either by debating ideas or by dividing the task into sub-tasks based on their individual capabilities, the overall performance can be significantly improved. For example, one agent might be excellent at generating creative content, while another excels at factual accuracy. By collaborating, they can produce a more comprehensive and well-rounded output than either could achieve alone. This collaborative approach allows for a more thorough exploration of solutions and can lead to more nuanced and higher-quality results.

8. What is the overall significance of adopting agentic workflows for LLM-driven systems?

The adoption of agentic workflows signifies a shift towards making LLM-driven systems more adaptive, autonomous, and capable of handling complex, multi-step tasks with a level of sophistication closer to human problem-solving. By incorporating patterns like reflection, tool use, planning, and multi-agent collaboration, LLMs can overcome limitations associated with single-pass generation and reliance solely on internal knowledge. This leads to improved performance in various applications, including coding, text generation, and decision-making. Ultimately, agentic workflows pave the way for LLMs to tackle more intricate and real-world problems effectively, unlocking new possibilities for AI applications.