# LLMs as Parts of Systems

DevFeed: [LLMs as Parts of Systems](<https://devfeed.tech/articles/llms-as-parts-of-systems-12575.md>)

Original publisher: [Read original article](<http://brooker.co.za/blog/2025/08/12/llms-as-components.html>)

Author: Marc Brooker

Published: 2025-08-12T00:00:00Z

Content type: opinion

Language: en

Sources: [Marc Brooker's Blog](<https://devfeed.tech/sources/marc-brooker-s-blog.md>), [Marc Brooker's Blog](<https://devfeed.tech/sources/marc-brooker-s-blog-2.md>)

Topics: [Large Language Model](<https://devfeed.tech/topics/llm.md>), [systems](<https://devfeed.tech/topics/systems.md>), [AI Development](<https://devfeed.tech/topics/ai-development.md>), [Automated reasoning](<https://devfeed.tech/topics/automated-reasoning.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [Kiro](<https://devfeed.tech/topics/kiro.md>), [Amazon Bedrock](<https://devfeed.tech/topics/amazon-bedrock.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [ai-agent](<https://devfeed.tech/tags/ai-agent.md>), [amazon-bedrock](<https://devfeed.tech/tags/amazon-bedrock.md>), [automated-reasoning](<https://devfeed.tech/tags/automated-reasoning.md>), [kiro](<https://devfeed.tech/tags/kiro.md>), [llms](<https://devfeed.tech/tags/llms.md>), [reasoning](<https://devfeed.tech/tags/reasoning.md>), [systems](<https://devfeed.tech/tags/systems.md>)

## AI overview

The article argues that LLMs are most powerful when combined with other system components, such as code interpreters, databases, browsers, algorithms, and SMT solvers. These combinations can solve problems that LLMs alone cannot and can deliver substantially better speed and cost efficiency. Amazon Bedrock's Automated Reasoning Checks illustrate this approach by using LLMs to extract rules and facts, then applying an SMT solver to verify logical consistency.

## Source excerpt

LLMs as Parts of Systems Towers of Hanoi is a boring game, anyway. Over on the Kiro blog, I wrote a post about Kiro and the future of AI spec-driven software development, looking at where I think the space of AI-agent-powered development tools is going. In that post, I made a bit of cheeky oblique reference to a topic I think is super important. I asked Kiro to build a Towers of Hanoi game. It's an oblique reference to Apple's The Illusion of Thinking paper, and the discourse that followed it. The question of whether LLMs can scalably play Towers of Hanoi is an interesting theoretically and scientifically, but not the most important question. The more important one is can systems built with LLMs play these games?. By picking me Towers of Hanoi in that other post, I was pointing out that the answer is clearly yes. And has been for several LLM generations. As a system builder, I'm much more interested in what systems of LLMs and tools can do together. LLMs and code interpreters. LLMs and databases. LLMs and browsers. LLMs and SMT solvers. These systems can do things, today, that LLMs alone simply can't, and will never be able to do. More importantly, they can do things today orders of magnitude more cheaply and quickly than LLMs can, even in the case where they can do the same things. You know, this kind of thing: > Generate a python snippet that counts the number of rs in a string. def count_rs(input_string): return input_string.lower().count('r') Trivial? Yes. But I've now created a system that that can solve problems that this LLM can't. A better LLM can, but at about six orders of magnitude higher cost per example. Systems, fundamentally, are more than the sum of their components. A good system can do things that no component can do alone. The trivial example is trivial, but you can imagine how that power could extend to being able to use decades of progress in algorithms. And not only count, but much more powerful things like SMT solvers, or ILP approximation, or