RLM’s design approach is not completely sound, especially in light of Brooks’ Law

[D7x7z49]

Hey @Hmbown, thank you for building DeepSeek-TUI and sharing the RLM idea. I really like it.

I have a small concern. When RLM calls many child agents in parallel, the answers they bring back may not agree with each other. Some answers could confuse the root agent or give it wrong hints. More parallel calls can make this worse. The cache hit rate may drop. And when you merge many results, agents might misunderstand each other's meaning. This feels like Brooks' Law.

I made a simple tool to think about this problem. It is called Issue Tree. It can label each sub-call as Task, Question, Condition, or None. The None type means the message has no real information. It is just coordination noise. This could help measure how much misunderstanding happens inside RLM.

You can find the idea here. Issue Tree at https://github.com/D7x7z49/llm-context-idea/blob/main/idea/issue-tree.md. Forgetful Society at https://github.com/D7x7z49/llm-context-idea/blob/main/idea/forgetful-society.md.

I wonder if you have seen this in practice. Could a metric like this help tune RLM's parallelism? I am very curious to hear your thoughts. This is just an open discussion. Thanks again!

[Hmbown]

Wonderful analysis - I have been unhappy with the RLM setup and the results I've been getting and this gives me a great head start!!

[Hmbown]

I think part of the issue may be the multiple sub agents / using RLM for a horizontal approach when it's meant for more vertical analysis.So I am thinking about just removing the multiple sub agent part and focus it back solely on one chunk before adding more back.

[D7x7z49]

Wonderful analysis - I have been unhappy with the RLM setup and the results I've been getting and this gives me a great head start!!

see: https://github.com/D7x7z49/llm-context-idea/blob/main/idea/lazy-tool-evaluation.md
see: https://github.com/D7x7z49/llm-context-idea/blob/main/idea/script-as-tool.md

see: https://github.com/D7x7z49/waf/blob/main/SYNTAX.ebnf

These should give you some ideas.

[D7x7z49]

@Hmbown

Your name and Shannon Labs caught my eye. Is the name a nod to Mr. Shannon, the father of information theory? I have deep respect for his work.

If I am thinking of the right Shannon, I would be grateful if you could look at a discussion here,
anomalyco/opencode#23628

The core of that issue comes from a conjecture I wrote,
https://github.com/D7x7z49/llm-context-idea/blob/main/idea/square-root-boundary.md

The starting point is simple. A token sequence S_N exists. I assume another token sequence S_k can match it in meaning. I do not look for the best S_k. I look for the cliff. How small can k get before no S_k can possibly keep the meaning? Pure information theory. Encode N states with k symbols. When k is too small, failure is certain.

Given your background, your perspective on this would mean a lot to me. I would be glad to hear your honest take.

This is also connected to our earlier chat in #1301. Some of the ideas there about Brooks' Law and agent coordination touch the same core questions.