In this post I will look at the different types of agentic systems. AI is moving fast and it is easy to become confused with the constantly evolving technologies. Let’s start!
There are two main types
1. Workflows
2. Agents
Workflows
A workflow is a series of steps that follow a predefined, rigid path. Predictability is high because you know exactly what the system will do, but this comes at the expense of flexibility. If one steps breaks then the whole process if likely to fail. There are five main types of workflow.
Prompt Chaining
This is probably one of the most common types of workflow out there. Given an input the LLM(Large Language Model) carries out a task and optionally hands the results to some code which will transform or clean the results before passing to another LLM, then these results can be passed to the next LLM in the chain and so on…
The key is that you are breaking down a complicated single task into smaller, manageable steps.
Routing
The routing workflow is where an initial router LLM analyzes and categorizes an incoming query and directs it to the most appropriate specialized sub-task.
Parallelization
The input is passed to a coordinator (code) that breaks the task into independent pieces. These run simultaneously across multiple LLMs. The last sub-task is not an LLM but some code that will take the results and aggregate them. This is best for speed and processing large volumes of data.
Orchestrator/worker
Here, an LLM acts as the manager. It dynamically decides which sub-tasks are needed and assigns them to “workers”. The Orchestrator then synthesizes the various results into a final response. This is more flexible than standard parallelization because the “manager” adapts to the complexity of the query.
Evaluator
In this workflow you have two LLMs in a feedback loop. One is the generator and one in the evaluator. The first LLM takes the initial user prompt and creates a draft. The second one reviews the draft against a set of given criteria and provides detailed feedback. The generator receives the feedback and produces a second version. This continues until the evaluator produces a pass or has hit it’s “max loops” limit. Without a “max loops” limit, an Evaluator and Generator can sometimes get stuck in an infinite loop (and burn your API budget!!).
Agents
Unlike workflows, Agents use a reasoning loop to determine their own path. They are characterized by their ability to use Tools—like searching the web or executing code to solve open-ended problems.
The LLM gets to choose it’s own design and plot it’s own path to choose how it will solve the problem. This power makes them very powerful but less predictable.
What are the drawbacks?
- Unpredictable path – you do not know which order the sub-tasks will be run but tools are used to give an agent boundaries
- What quality will the output be?
- Unpredictable costs, you don’t know how long it will take to run
Mitigations
- Monitoring – it is essential have the visibility to understand what interactions are going on
- Guardrails – protection to ensure models are doing what they should be doing, safely, consistently and within the given boundaries
Conclusion
In the current landscape of AI engineering, Workflows remain the most popular choice for production-grade applications. This is because businesses value reliability and cost-control. Patterns like Prompt Chaining and Routing allow developers to build systems that are fast, explainable, and easy to debug. If you are building a customer support bot or an automated report generator, a structured workflow is usually your best bet.
However, the industry is rapidly shifting toward Agents and Evaluator-Optimizer loops for high-stakes or creative tasks. While more “expensive” in terms of compute and time, these systems provide a level of quality and autonomy that simple chains cannot match. They are becoming the standard for coding assistants, research tools, and complex problem-solving.
The Rule of Thumb
- Use Workflows when the process is well-defined and you need 100% consistency.
- Use Agents when the task is open-ended and the path to the solution is too complex to map out by hand.