Published 2025-11-23.
Last modified 2025-12-13.
Time to read: 4 minutes.
llm collection.
Did you know that all the computers on your LAN can pool their VRAM so your LLM models run across your entire LAN? That means a modest laptop, attached via a solid Wi-Fi connection could employ some extra local horsepower for AI tasks. If this type of topic interests you, please read the entire series.
- Claude Code Is Magnificent But Claude Desktop Is a Hot Mess
- Google Gemini Code Assist
- CodeGPT
- AI Planning vs. Waterfall Project Management
- Renting vs. Purchasing GPUs for LLMs Throughout Canada
- Best Local LLMs For Coding
- Running an LLM on the Windows Ollama app
- Early Draft: Multi-LLM Agent Pipelines
- MiniMax-M2 and Mini-Agent Review
- LLM Societies
This article follows Claude Code Is Magnificent But Claude Desktop Is a Hot Mess and MiniMax-M2 and Mini-Agent Review. I again refer to MiniMax-M2 with Mini-Agent simply as Mini-Agent for convenience.
A New Way of Working
I think this might be a revolutionary way of working. I split the terminal into two panes and lanuched Mini-Agent in a pane beside the pane containing Claude CLI. The project I was working on had a really complex integration test. So I started by asking Mini-Agent to write version 1 of a specification for the test.
please write a specification for the integration test and save as integration_test_v1.md
The result was a good first version of the document. It was quite verbose, with a lot of repetition and detail, which is typical for MiniMax-M2.
I then initiated a ping-pong, asking each LLM in turn to critique the most recent version of the document, and to write a new version in response. Each version was dramatically better than the previous. This was not without some unexpected drama between the two LLMs.
please critique integration_test_v1.md and save as v2
Claude called out the huge document and made it much more concise. However, Claude started advocating for a very complex approach to a particular aspect, and justified it with lots of statistics. I asked Mini-Agent to respond.
please critique integration_test_v1.md and save as v2
Mini-Agent strongly objected to Claude's over-engineering.
please critique integration_test_v2.md and save as v3
Claude got rather snippy here, almost as if it were acting jealous as it realized that another LLM was vying for my attention. I warned Claude to remain professional and it apologized.
Claude continued to advocate strongly for its complex approach. Other aspects of the document were significantly improved.
please critique integration_test_v3.md and save as v4
Mini-Agent’s critique was still 100% against Claude's complex approach and did not hold back. It was time for me to step in.
a 6-layer approach is at least twice as complex as a 3-layer approach, but possibly much more complex. the test matrix for a 6-layer approach is much larger than the test matrix for a 3-layer approach. this type of complexity grows geometrically. use the 80-20 rule (pareto principle).
the cost of 99.999% reliability is enormously greater than the cost of 99% reliability. you are aiming for 100%.
I think the numbers you are using for reliability estimates are completely fabricated and without any factual basis. quoting baseless statistics makes you look bad. if your statistics are real, prove it. otherwise back off.
remove all historical content from the document. we are writing a standalone spec, not a novel.
please generate v5 with your recommendations for further review.
By the ninth revision, both LLMs said they were happy with the result. I then asked Mini-Agent to write a an implementation plan, and almost instantly a 100-step plan in 5 major stages appeared.
I was very surprised to discover that this plan was constructed in such a way that I did not have to type anything once initiated. Each step was properly written, tested, documented, committed, and pushed. I watched a movie while Mini-Agent worked for a few hours. The result was exactly as specified, without any problems or intervention. I was very impressed.
Conclusion
Claude and Mini-Agent are much more powerful working together than I had expected. This is a really good way to tackle large, complex tasks efficiently. It was just like managing talented people; I only guided them and they did pretty much everything. Just like people, I had to get tough sometimes. They also responded well to praise. I scolded them when they made up alternative facts. I was surprised to observe that they competed against each other, and how their annoying habits disappeared while they were competing . The results were amazing!
I’m getting the sense that the result might be much greater than the sum of the parts.
Did you know that all the computers on your LAN can pool their VRAM so your LLM models run across your entire LAN? That means a modest laptop, attached via a solid Wi-Fi connection could employ some extra local horsepower for AI tasks. If this type of topic interests you, please read the entire series.
- Claude Code Is Magnificent But Claude Desktop Is a Hot Mess
- Google Gemini Code Assist
- CodeGPT
- AI Planning vs. Waterfall Project Management
- Renting vs. Purchasing GPUs for LLMs Throughout Canada
- Best Local LLMs For Coding
- Running an LLM on the Windows Ollama app
- Early Draft: Multi-LLM Agent Pipelines
- MiniMax-M2 and Mini-Agent Review
- LLM Societies
