I have the first tier subscription to 4 models - claude, chatgpt, perplexity and grok and also have an openclaw setup. openclaw monitors all my daily wearable data as well as workouts on my carol bike. I have recently found out about buzz.xyz which allows models to debate with each other - plan on downloading that this weekend
3 Likes
Here: Claude, ChatGPT, GEmini, Grok, perplexity. Openrouter if I want to try something different.
But presently there is a duopoly: Claude and GPT are by far the best ones.
Gemni is lagging behind and Grok, although good in some tasks, is way too pricey for what it gives (the optimization has started to make it dedicated mainly to developers, rockets and electrical cars ).
Recent Kimi3 exhibited extraordinary capabilities for an open source model, but subscriptions are closed and it is available only through API keys.
Buzz.xyz= good tip. GPT5.6 Sol debating with opus5 or Fable5 is going to be epic, looking forward to it.
1 Like
Alex Finn describes buzz as an adversarial agents arena. He uses it for coding, but many things designed for coding turned out to be good for non coders as well (i.e.: Claude code).
@Paul
Let us know pls about your installation and setup. I use no Openclaw or hermes presently but the setup sound very useful for any critical discussions, I would like to apply it using my main nodes: GPT, Claude, Gemini, as discussing experts, led by me as main coordinator of the brainstorming.
Consulted Claude Code, Buzz seems to be too specific for coders. I am going to try Karpathy’s Council of LLMs maybe. Waiting for reviews on buzz.
The five hard truths:
-
There is no “add agent” button. Each agent is a separate process you launch on your own laptop, with its own Nostr private key (
nsec1…), its own env vars, its own API key. Four agents = four terminal windows humming on the HP Pavilion. Close the lid, the panel dies. - Resource load. Each agent spawns its own MCP subprocess. The docs recommend starting at N=2; we want 4. On a 155H that’s survivable but not comfortable, especially with reasoning models chewing on long contexts.
-
@mention-gating vs. debate. By default agents only speak when @mentioned — great for cost, useless for a free-flowing panel. Forum mode (
--no-mention-filter) makes them respond to everything, which is what we want and is exactly how you get four models talking past each other in an unbounded, billable loop. No published loop-prevention beyond turn-duration caps. - Cost is not linear. In a group chat every agent re-ingests the whole growing thread on every turn. Four agents × a two-hour session = a genuinely surprising API bill. Rough order: a single meaty brainstorm session runs single-digit to low-double-digit dollars; sloppy forum-mode looping can multiply that fast.
- Wrong instincts. These are coding agents. Pointed at “let’s brainstorm my consultancy positioning,” they’ll behave like coding agents in a trench coat unless we invest real work in system prompts and personas.
Verdict on Buzz: feasible, roughly 4–8 hours of setup for someone comfortable in a terminal, ongoing babysitting, and it is squarely a developer product. The value it adds over a simpler panel — persistent history, agents that can act on files and repos, cryptographic audit trail — is value we don’t currently need for brainstorming.
3. The alternative that actually matches the goal
Karpathy’s llm-council (~23k stars) does precisely what you described: one question goes to N models in parallel, each model then reviews and ranks the others’ answers anonymised, and a chairman model synthesises the final take. One OpenRouter key, one web app, no Docker, no four terminals, no per-provider accounts. There are hosted versions too.
What you lose: it’s turn-based Q&A, not a persistent room, and the models can’t do things — only argue.
Honest read: Council gives us ~80% of the intellectual value at ~10% of the setup cost. Buzz earns its complexity only once we want agents that act — open repos, edit the strategy doc, run workflows — and only once the Gemini/Grok bridges are proven by someone other than us.
1 Like
I just used Karpahty’s LLM council for a technical question based on Quantitative Risk Analysis, one of my interests. A little work to set it up, but, honestly, Claude Code made it all.
Basically, a debate among Opus5, GPT5.6, Grok 4.5, KimiK3 and Gemini 3-1 Pro.
I was awed. It works better in topics where there is not much consensus (well-suited in health & longevity!).
I wonder what may be a question of general interest here.
2 Likes
How did you set this up?
1 Like
With the help of Claude Code. It did all the setup, I only clicked ‘YES’ at certain stages.
1 Like
I use Claude (pro) near-daily and have since 2025. I employ multiple log and analysis files that are updated and which I load into new instances to keep the usage meter under control and get fresh perspectives.
In addition, I load a guidelines file that is extensive and detailed about how to interact with me, present data, organize ongoing hypotheses and questions, and stay within guardrails. I have a whole section about how the instance should self-reflect to most effectively reason, what to watch out for, and when it is most likely to be wrong.
This is not a panacea. I have to also think critically and ask and re-ask and challenge, even when I’m asking the instance to do its own challenging.
I thought about uploading my guidelines file for folks here, but this thread is already long. If anyone wants to see it, direct message me.
[THIS MESSAGE entirely human-generated, tho note from the brain of a post-migraine human 
.]
4 Likes
After trying to set up Buzz I found it rather challenging to do what I wanted but have found the greatest Agentic app - GrokBot - This is super easy to set up and with the built in computer for each agent i have it logged into various accounts. Very interesting to watch it work and the agents discuss ideas with each other
2 Likes
A new and good resource I’m using to check on some new registrations and posts here (from suspicious people/ accounts that may be spam). You may find it valuable when you’re wondering if a given piece of text is AI generated or not:
Reviews:
1 Like