Feedback is all you need

When building multi-AI agent systems, you’ll sometimes find that models struggle to carry out instructions with complex requirements.

There are several solutions to this problem. The most common ones include improving your prompts, swapping in a more capable LLM, and—the topic I want to discuss today—adding a reflection loop.

I wanted to explore this subject not only to share insights on building better multi-AI agent systems, but also because I noticed something striking: it closely mirrors how humans operate.

Improving prompts is analogous to giving clearer, more understandable instructions to a person. Naturally, someone will perform better when they’re given more detailed requirements and guidance on how to achieve them.

However, if the worker lacks the necessary skills or background knowledge, they’ll still struggle to get the job done.

The second approach—replacing the model with a more capable one—is like swapping in a more talented worker.

Someone with greater expertise and a solid foundation of knowledge will undoubtedly produce better results.

But here’s the catch: talented workers are hard to find, and they demand higher compensation.

The same applies to LLMs. More capable models cost more, and they hit rate limits more frequently.

The final approach is to add a critic agent that enables agents to exchange feedback and reflect on their own shortcomings.

In human terms, imagine a team collaborating—sharing ideas, making judgments together, and giving each other feedback.

What’s remarkable is that this reflection loop approach often outperforms simply upgrading to a more capable model, both in terms of performance and cost.

We see the same phenomenon when people work together.

Sure, sometimes a single exceptional individual can drive great results. But more often, even when each of us has our own weaknesses, we collaborate, grow through mutual feedback, and create something wonderful together.

So whether we’re architecting multi-AI agent systems or playing our part as members of society, perhaps we should never forget the power of collaboration and feedback—and keep moving forward together.