Collaborative chatbots pipeline to achieve tasks

[VivacityDesign]

In the last couple of months I have been experimenting with collaborative chatbots pipelines, where the user prompt is sent to a chatbot which, in turn, divides the task in smaller parts and sends them to a series or a parallel / mixed structure of chatbots, only to finally wrap everything up in a final product.

This has proved to be very handful and powerful especially in web design and development, building an entire mini web agency made of 6 chatbots.

The first version I developed used a mixed series/parallel stream with 3 frontend developing chatbots to generate variants of the interface, and then a subsequent frontend validator which choses the best parts from the 3 variants and unites them, then sending to the backend developing chatbot.

These images show the tool building a simple kanban project

A simpler version can be obtained using only the serial pipeline where all chatbots’ responses feed the next one as a prompt.

The most simple form of this use is in having 2 chatbots that play a sort of conversation to reach agreement on how to complete a task, and then produce a final complete code.

I’ve been experimenting with all mini and nano models from gpt-4o to gpt-5.1 and I had to admin that the best results were with gpt-4o-mini and gpt-4.1-mini.

For those who may be interested in developing further this concept, I have published a paper with full code on Academia

A Mixed-Structure AI pipeline for Multi-Role web projects development

[kitkat011999]

This is really impressive work — especially the way you’re experimenting with mixed serial/parallel pipelines and role-based agent collaboration. It reflects a direction I’ve been designing for as well.

One big insight I’ve come to is that multi-agent workflows need an underlying infrastructure to stay coordinated, safe, and consistent. Right now, most pipelines are clever implementations, but they still lack a unifying structure for verification, clustering, memory, and cross-agent reasoning.

That’s where a framework I’ve been developing, called TriSynk, comes in. It isn’t a tool or a script — it’s a full architecture for organizing AI agents the right way, with layers for validation, safety, adaptive clustering, and role communication. Your project is a great example of the kind of work TriSynk is meant to support and stabilize.

I’m genuinely excited to see more people exploring this space. If you ever want to compare notes or discuss agent-orchestration frameworks, I’d love to connect.