DEV Community: Ricky FF

Introducing Minotauris: Why One AI Agent Shouldn’t Do Everything

Ricky FF — Sat, 18 Jul 2026 02:13:51 +0000

Most AI agent harnesses follow a similar structure.

One model receives an objective, creates a plan, operates tools, reviews the outcome, and decides whether its own work was successful.

That structure becomes fragile when the task grows beyond a few tool calls.

A mistake in planning affects execution. The same context that produced the mistake is then responsible for detecting it. Tool calls become repetitive, model usage increases, and the user still has to supervise the entire process.

Minotauris takes a different approach.

A structured AI team

Minotauris separates responsibility across four parts.

Leader

The Leader maintains the objective and sets direction without performing every execution step.

Managers

Managers divide work, coordinate Workers, and preserve the context required to complete the larger objective.

Workers

Workers receive focused responsibilities and execute tasks across the browser, desktop, code, files, and tools.

Assembly

The Assembly gives the system a dedicated place to challenge the plan, inspect assumptions, and consider possible outcomes before execution.

The point is not to create more agents for the sake of having more agents.

The point is role separation.

Connected to the computer

Minotauris currently operates as a Windows beta with browser use, screen and computer control, scheduled tasks, remote access, coding, and general task execution.

The larger question behind the project is simple:

Can AI systems become more reliable and efficient by organizing intelligence as a coordinated team instead of placing every responsibility inside one model loop?

Demo and product:
https://www.minotauris.app/

Why AI Agents Need Better Team Architecture, Not Just Better Models

Ricky FF — Sat, 11 Jul 2026 22:50:34 +0000

A more capable model does not automatically create a more reliable agent system.

Most current harnesses ask one model to plan, execute, review, and recover inside the same loop. Minotauris explores a different architecture: a Leader, Managers, Workers, and an Assembly with separate responsibilities.

This article examines why role separation, critique before execution, and coordinated computer control may matter as AI moves beyond chat.

I Vibe-Coded a "Cursor for Writers" on a $7 budget (Because Chat kills the flow)

Ricky FF — Wed, 28 Jan 2026 06:09:16 +0000

We talk a lot about "Vibe Coding"—using AI to stay in the flow state while building apps.

But what about "Vibe Writing"?

Every time I try to write a novel with ChatGPT, the vibe dies instantly.

I have to stop writing.

I have to explain the plot again.

I have to copy-paste text.

The Flow is gone.

I’m 16. I don't have VC money. But I have $7 and some Supabase/Groq credits. So I built Minotauris.

It’s an Agentic IDE for fiction.
⚡ The Vibe Check (How it works)

Instead of chatting, you Command.

Stay in the Editor: You never leave the writing interface.

Async Agents: You type /refactor chapter 4 for pacing, and the agent runs in the background. It reads your Lore, checks your tone, and edits the text while you keep writing Chapter 5.

No Context Amnesia: I indexed the "Story Bible" so the AI actually remembers the villain's motivation from 20 chapters ago.

🛠️ The "Broke" Stack

I vibe-coded the whole thing in 2 weeks:

Brain: deepseek-r1 (via Groq) for the hard logic.

Speed: llama-3-8b (via Groq) for the fast drafts.

Routing: Supabase Edge Functions (to throttle heavy users so I don't go bankrupt).

I’m looking for people who want to test the "Command Mode" workflow. If you’re tired of the Chat Box killing your creative vibe, let’s talk.

Waitlist / Alpha:
minotauris

(P.S. If you have any tips on optimizing RAG for 100k+ token stories without melting my credit card, let me know in the comments 👇)

I’m 16, broke, and I built a "Cursor for Writers" using Next.js + Groq ($7 budget)

Ricky FF — Wed, 28 Jan 2026 05:52:03 +0000

The "Chat" interface is killing our creativity.

We have Cursor for coding. It knows our repo, understands our dependencies, and writes code in context.

But for writing novels? We are still stuck copy-pasting text into a "Chat" box that forgets the main character's name every 4,000 words.

I couldn't afford a $30/month subscription to enterprise tools. So, I spent my last $7 building my own Agentic IDE on the Groq LPU.

Here is the tech stack, the architecture, and how I solved the "Context Amnesia" problem on a high school student's budget.
🏗️ The Problem: "Context Amnesia"

Standard LLMs have a "Slide Window" memory. If you are writing Chapter 20, the model has likely forgotten the foreshadowing you planted in Chapter 1.

Most "AI Writers" are just wrappers around GPT-4. They treat your novel like a conversation, not a project.

I needed an Agent, not a chatbot.
🛠️ The "Broke" Tech Stack ($7/mo)

I call it Minotauris. It’s built to run lean but act smart.

Frontend: Next.js 15 (App Router) + Tailwind CSS.

Backend: Supabase (Auth + Database).

The Brain (Smart): deepseek-r1-distill-llama-70b (via Groq) for reasoning/plotting.

The Muscle (Fast): llama-3.1-8b-instant (via Groq) for prose/drafting.

The Orchestrator: A custom TypeScript router that decides which model to use.

🧠 The "Smart Router" (How I save money)

Since I'm 16 and not VC-backed, I can't let users burn $100 in API credits. I built a Traffic Router in Supabase Edge Functions:
TypeScript

// The logic is simple:
// 1. If the task is "Planning" -> Use Expensive Model (70B)
// 2. If the task is "Drafting" -> Use Cheap Model (8B)
// 3. If User > Monthly Limit -> Enforce "Slow Pool" (Artifical Delay)
![ ](https://dev-to-uploads.s3.amazonaws.com/uploads/articles/427v5m4p77pcilttljlg.png)

This lets me offer "Unlimited" basic drafting because Llama 8B on Groq is practically free (~$0.05 per million tokens).
⚡ The Agentic Workflow

Minotauris isn't a chat. It’s an Asynchronous Command Center.

The Lore Index: I vector-embed the user's "Story Bible" (Characters, Locations, Rules).

The Command: User types: "Refactor Chapter 4 to be more suspenseful."

The Sub-Agent: The AI pulls the specific lore needed for Chapter 4, checks the tone of Chapter 1, and rewrites the text in the background.

I also built a Mobile Command Mode. You can send commands from your phone while walking the dog, and the cloud agent executes the heavy lifting on the server.
"Demon Mode" Development

I built the MVP in 2 weeks using v0 for the UI and Cursor for the backend logic.

I’m currently running a closed Alpha because my wallet can't handle the whole internet yet. But if you are a developer who writes (or a writer who devs), I’d love your feedback on the architecture.

The Waitlist is here: 👉 minotauris.app

Let me know in the comments: How are you handling long-context in your own AI apps? RAG? Long-context caching? I need optimization tips before I go broke lol.