When writing this post, I just finished the first day of SWITCH (Singapore Week of Innovation and Technology). It is one of the biggest tech conferences in Singapore where many tech companies, startups, VCs from other countries, and Singapore come together to learn, and network.
I am at the AIMX (pronounced as AI-MAX) stage and was listening to AI thought leaders and experts. I very quickly concluded that agentic workflow is going to be the AI theme for 2025. If you are interested in understanding a bit more about agentic workflow, I will be providing two resources at the end of the issue, so look out for them. :)
I want to discuss my thoughts on how to implement a successful agentic workflow.
There are two layers to overcome if you want multiple successful agentic workflows that allow for customization and maintenance. What are the two layers? Let me explain the first.
First Layer: Elemental Tasks
Firstly any agentic workflow or big task, such as fixing a leaking water pipe, will have to be broken into its elemental tasks. If you know chemistry, elements are substances that cannot be further broken down unless it’s into protons, neutrons, and electrons. Successful breakdown of any big tasks into elemental tasks will be important as it has a downstream impact later on, and not forgetting maintenance cost as well. But the breakdown into elemental tasks will be the first layer of it for sure.
Second Layer: Selection, Combination, and Orchestration
Let’s use an example to explain this better. Let us go back to the above example of fixing a leaking water pipe. You can imagine there is a toolbox of agents/tools that perform and get specific results/impact.
So how do we fix that leaking pipe? We will first have to map out the steps that are needed to achieve to final result, which is a non-leaking pipe. Once the steps are mapped out, we will need to select the tools/agents that are needed for each step, some steps might need multiple agents which compound results is what we want.
The mapped-out steps will show the combination of each agent or agents, depending on the steps. For example, Step A may need a wrench first, followed by Step B which may need a metal nut together with a spanner, and Step C may need a washer and tape. These tools mentioned, like the spanner, metal nut, washer, tape, etc can be seen as agents.
The mapped-out steps not only dictate the selection and combination but also determine the orchestration as well. Going back to the example, the spanner needs to be used to ensure the metal nut is tightened first, then we can move to Step C, to use the washer and the tape. Orchestration can be seen as the timing needed for each agent, and then moving on to the next set of agents for the next stage.
The difference between combinations and orchestration is combination says how the agents are combined together together, the sequence, whereas, orchestration is how long each stage is, or to what conclusion before the workflow can move to the next stage.
You can see from here that if the elemental tasks are not well-defined, it can impact all three of them, selection, combination, and orchestration because these agents cannot fit into each other nicely to form a workflow. It is like doing a jigsaw puzzle where each of the puzzle pieces selected must come together EXACTLY to form a nice picture. Thus defining each agent is going to be very important for agentic workflow to work and to be able to scale and maintain easily.
I am looking forward to the age of AI agents though as it will further give us insights into how humans work vs how machines work and hopefully result in how humans and machines can work together to increase the impact created to society.
I still have a few more thoughts on Agentic Workflow that I will share as we go along and I look forward to your feedback and comments as well. PM me on LinkedIn or just state it in the comments below. :)
Support my work, share my newsletter with your community, and network! :) Feeling generous? Drop me some “books” please! It will help me to get more materials for my learning! :)
Love it and it can be turned into a framework that would help translate a topic one wants to address with agentic AI into specifications for development.
Let's call it the "Koo Agentic AI Specification framework" (KAAISF) : Give me a problem you are trying to solve that requires different skills along the way. I will then define the 2 layers as per your article:
Layer 1: Elemental Tasks
- Break down complex tasks into indivisible components
- Precise definition is critical for downstream effectiveness
- Impacts scalability and maintenance costs
- Similar to chemical elements that cannot be further reduced
Layer 2: Selection, Combination, and Orchestration
- Selection: Choose appropriate agents/tools for each step
- Combination: Define how multiple agents work together for each task
- Orchestration: Determine timing and sequencing of agent actions
- Steps must fit together precisely like puzzle pieces, to use your analogy
We would then use the various agentic AI frameworks available today to turn the topic into reality.
1. Foundation Layer Implementation -> LangChain is particularly strong for implementing the foundation layer; Autogen excels at implementing base agent architecture
2. Orchestration Layer Implementation -> LangGraph is excellent for implementing workflow orchestration; Crew.AI is particularly strong for agent coordination
3. Intelligence Layer Implementation -> we would combine multiple frameworks for advanced capabilities