Your Shitty Team Of Specialized Agents

What's the purpose of splitting the work between different specialized agents?

Are you following a tutorial without thinking it through, or are you actually solving a real problem you encountered?

When trying to solve the “scaffold any simple NextJS frontend in less than 10 sec”, one of my first attempt was of course to use CrewAI.

I would have a team of agents assigned with Ollama 120b on Groq, meaning 1,200 TPS for the sake of god.

For those who always wondered what CrewAI is actually doing under the hood, I implemented a visualization tool you can find on this repo.

Yet, it would take several MINUTES to generate a simple project.

Why?

💡 Despite Groq's 1,200 tokens/second speed, execution feels slow because:Each task runs through 15–50 sequential LLM round-trips (agentic tool-calling loop)CrewAI adds 3 sequential agents per task (Planner → Developer → Reviewer)Batch execution runs tasks one-by-one, even when they're independentResult: tasks take minutes instead of seconds

Yeah. That’s what you are doing to yourself. All those useless messages exchanged, for what?

Despite some attempt to optimize the process, you need to zoom out and ask yourself the real added value.

Why do we treat LLM like people with one job and one expertise?

This does not even apply to myself since as a co-founder, I occupied those position intermittently in the past: Software development, Software architecture, Full-stack development, Code review, Technical estimation, DevOps and deployment, Machine learning engineering, AI model integration, Quality assurance, Testing and debugging, User acceptance testing, Product management, Product discovery, Roadmap prioritization, User research, Customer education, UX collaboration, Design support, Project management, Delivery management, Team coordination, Risk management, Business analysis, Proposal review, Pricing and estimation, Strategic planning, Vision definition, Account management, Client relationship management, Marketing strategy, Go-to-market strategy, Content writing, Blog writing, Social media management, Analytics and metrics analysis, Data-driven decision making, Leadership, Recruitment and interviewing, Customer support, Financial operations, Accounting management, Operations management.

And yeah, I have a senior level in: iOS development, Android development, React web development, Flutter, Expo, etc…

So, if one person can do all of this, why should we split this work for an agent... like it was too much?

In most cases, you probably need only one super agent, with a disposable team to spin up (very fast executors, and possibly almost free), to get your job done faster and cheaper.

You just need to think about it.

This also reflects the direction of Warp’s new feature, Oz (note: a single run consumed 600 of the 2,000 free credits provided).

Stop copy pasting other people skills found on the web, create skills tailored to your needs.

Next question for your quest: what did Clawdbot truly brought to the world that was not already in it?

Spoiler alert: nothing

Small quizz

For those who still assume local LLMs aren’t viable options.

On a MacBook Pro M4 Max 64 GB.

At which speed can GLM 4.7 run locally?

At which speed can qwen-3-coder-30b run locally?

At which speed can qwen-3-coder-next run locally?

Apple to soon win the AI war with the M5 Ultra 1 TB, without participating the race.