22 small LLMs beat one big brain on control and cost
Most AI projects still chase bigger models.But the ROI breakthrough comes from small, specialized teams of LLMs.
When I design AI systems, I don't think in "one big brain".I think in mini-organizations.
Each agent has:• one task• one or several handovers• one clear role
Example from our work: processing tens of thousands of emails for a lawsuit.
We break it down into:• OCR extraction with 2 different models for quality checks• Models that identify where an email starts and ends• Models that identify the sender, timestamp, create summaries, and categorize• Quality management LLMs baked throughout
(In reality the project was much more complex, but this gives you the idea)
22 LLMs.22 clear jobs.
Here's what happens when we design systems like this:
- We see where things break
- We swap weak parts without touching the rest
- We use light models where frontier models aren't needed
More control.More explainability.Less cost.Less energy use.
Now a new layer enters the game.
Kimi K2.5 was just released and it's as powerful as proprietary frontier models.
The interesting part? It builds its own "mini organization" inside the model.
I give it a task.The model creates an agent swarm:• up to 100 small agents• working in parallel• each with a specific role
Results in some tests:• up to 4.5x faster than running a single model sequentially• more complex work in one go• no manual role design needed
Sounds great.But here's the tension.
When the model builds its own swarm, I gain speed but lose direct control.
- I see less of what happens inside
- I depend more on the orchestration logic
- New hidden failure modes emerge between agents
In regulated environments or mission-critical processes, this is a problem.
Don't get me wrong - this model is fantastic and proves once again that open-source is just as powerful as Claude, GPT and Gemini. I believe open-source will dominate performance for most use cases going forward.
But it's a design choice.
Build your own agent system when:• You need audit trails (legal, compliance, finance)• Errors have direct business costs (customer communication, invoicing)• Quality must be consistent across thousands of cases
Let the model build its swarm when:• You're exploring new territory without clear process• Speed matters more than perfection• Human review happens before any action
Everyone talks about more powerful agents.Few talk about responsible architectures.
A system you can open up, debug, and evolve beats the most powerful black box in.
Size and power are impressive.When the goal is ROI, reliability wins.
Where would you place your bet - power or reliability?