AI Agents vs Workflows: What Changed (2026)

Añadido: 2026-05-20

[00:00:00]Before agents became a thing, we relied most of our work using workflows. There are different technologies that helped automate jobs, and all of them required a lot of engineering time and business analysis to create the system properly.

[00:00:14]LLMs started to emerge beyond a simple chatbot in ChatGPT, but they started to have bigger context windows and better reasonings. And LLMs now started to question how we can start changing the rigid way of doing things. Now, context window and reasoning are certainly useful, but what really changed the game when it comes to morphing LLMs into an agent was tool calling. Tool calls effectively changed the game since LLMs could now start using external tools that gave them the ability to start changing things outside of their own environment. And that's exactly around the time when we started to get agentic applications emerge like Cursor and Winsor. And the first real use case of an agent was directed at coding. Now, agents today can help deep research in ChatGPT and Manas. They help us troubleshoot and take calls, as well as build web applications. A great example is Lovable that can help build web apps, Perplexity that helps researchers and understand various sources from the web, Manas that helps with generic tasks that we just throw at them. All of these are use cases of an agent that underneath what characterizes an agent, unlike workflow, is that LLM that serves as a core decision maker that drives them.

[00:01:31]The agent is equipped with its own prompt, meaning they're given their own instruction, typically written in text, and their own context window that keeps track of information that is pertinent for the task. And comparing agents from, say, 2022 to agents today in 2026, we've gone through many iterations and changes. Back then, since context window was small, we heavily relied on prompt engineering, meaning how to best instruct an agent to actually get the job done. We later moved away from prompt engineering to context engineering as the context window grew, and we found that keeping the context organized had a better effect on getting things done. And the core difference here is what context window allows us to do things differently, which is time. As the context window grew, their ability to handle longer tasks also grew with it. You can see from this chart from Meter that AI is solving tasks at successful rates for tasks at different complexity in time. Agents today have shifted once again from context engineering to a much more mature system called agent harness that focuses more on the agent's environment that the agent is under rather than the specific mechanism that is given to help guide its decision and actions. As you can see, agent is an evolving practice and concept, and as the underlying technology continue to mature that gives them higher intelligence, bigger context window, and faster inference, or faster tokens per speed, all of these breakthroughs inherently changes what an agent is and what we can give the agent to get done.

[00:03:09]>> [music]