A Human-Centered Multi-LLM Workflow Inspired by CrewAI (Case Study)

Hi everyone,
I’d like to share a personal case study that might be relevant to those exploring human-in-the-loop or human-centered multi-agent workflows.

Why I’m sharing this here

Most multi-agent frameworks focus on planning, automation, and task efficiency.
My experience came from the opposite direction — a deeply human one:

What happens when someone under cognitive overload uses multiple LLMs as a support team to rebuild their thinking?

This became a manual but surprisingly effective multi-LLM “Externalized Brain” system, inspired by the idea of agent collaboration.

During a period of severe overlapping stress (300+ LCU), my cognition and judgment were impaired.
So instead of relying on a single model, I began assigning roles across three LLMs:

  • ChatGPT — structuring, integration, drafting

  • Gemini — logical critique, refutation, blind-spot detection

  • Copilot — third-party tone, audience perspective, ethical framing

By copying the same question across the models, each provided a different lens.
Together, they formed a three-perspective thinking team that helped stabilize my decision-making and reduce cognitive load.

This wasn’t automated at all — it was entirely conversation-driven — but the patterns mirrored many multi-agent concepts such as role specialization, cross-checking, productive friction, and iterative refinement.

I documented the workflow, the roles of each LLM, and how this “Externalized Brain” method supported cognitive recovery:

:link: **Case Study:**Externalized Brain: From Burnout to Recovery

Key elements covered in the write-up

  • Human-centered multi-agent collaboration

  • Conversational externalization of thinking

  • LLM role assignment and sequential reasoning

  • Using refutation and cross-model contrast to remove bias

  • Reducing cognitive load through distributed cognition

  • Connections to multi-agent research (Google, Meta, OpenAI, AWS)

Why this may interest the CrewAI community

Although my system is manual, the underlying logic overlaps with agent orchestration principles — especially for:

  • human-in-the-loop design

  • cognitive offloading with LLMs

  • multi-model role distribution

  • mental-state-adaptive workflows

If this resonates with your work or CrewAI’s direction, I’d love to discuss human-centered agent UX or how such workflows might evolve with proper automation.

Thank you for building tools that help inspire experiments like this.