Agent frameworks and workflow builders exist because the LLM model itself is not yet strong enough to be autonomous in completing the tasks.

It can be now. For those who are technical enough to do post training. And these post trained LLMs will render most agent frameworks and workflow tools obsolete.

Workflow Builders

We human users find comfort in using no code or low code tools to draw boxes on a blank canvas to create workflows and schematics. It feels great. It feels in control. But history has repeatly shown that this style of working quickly becoming obsoltele as technology progresses.

Let’s use a very recent example of this phenomenon as a case study.

Before the release of ChatGPT in late 2022, Gen AI was dominated by image generation models. It’s very common for hackers and builders to stitch together a bunch of image generation models in a workflow to generate the desired effects. Stable diffusion, image upscaling, LoRA, ControlNet, etc. The workflows are very customized and complex. And from here, ComfyUI was born in January 2023.

ComfyUI is an very advanced generative AI workflow tool that enables users to stitch together complex workflows involving many models, prompts, parameters, customizations to achieve the intended effect. Below is one of the image I find on Google. And this is not even a complex workflow. From this. many indie hackers built their buisiness on top of this.
alt text

The situation changed rapidly after ChatGPT launched its image generation capability in March of 2025. In mere two short years, model capabilities drastically improved. Creaters no longer need to stitch together complex workflows to achieve the same effects vs using ComfyUI.

We can now use one single prompt to edit images. Hell, we can even now generate a full short video using one single text prompt.

The model is the product.

alt text

Robotic Process Automation is not Agentic

Currently there are more than a few tools that masquerades and rebrands robotic process automation (RPA) workflow builders as ‘agentic’. There’s nothing agentic about these tools. Nada. Nil. Zip.

Don’t get me wrong. They are awesome tools for software engineering consultants to quickly stitch together some solution and sell to some non-technical customers to automate some tasks. More often than not, going from 0 to 1 captures majority of the value and there’s not enough value to scale from 1 to 100. The tiny white elephant that is burdensome to maintain but has enough value to be hanging onto.

Langflow, make, n8n are the perfect tools here, but they are absolutely not agentic.

alt text
alt text
alt text

Reinforcement Learning Agents

On the other hand, we are already witnessing full agentic behaviors in apps like ChatGPT, Claude Desktop, and Gemini App. Though they cannot do a very good job yet for some longer running tasks, they are absolutely amazing with their agentic capabiltiies including reasoning.

And some of us can achieve the same in narrower domains. Today. With proper system optimization and post-training. Without drawing boxes on canvas.

Conclusion

As AI/ML practioners, we can either choose to fight the last war or skate to where the puck is going. My preference is for the latter.

@article{
    leehanchung,
    author = {Lee, Hanchung},
    title = {No Code, Low Code, Full Code},
    year = {2025},
    month = {06},
    howpublished = {\url{https://leehanchung.github.io}},
    url = {https://leehanchung.github.io/blogs/2025/06/26/no-code-low-code-full-code/}
}