Gary Harrison

[Gary Harrison @Gary__Harrison]

Most multi-model tools work like this:

Ask Claude ✓
Ask GPT-5 ✓
Ask Gemini ✓
Pick any you like ✗

Get opinions, not consensus.

What if AI models debated and converged?

Thread 🧵 #AI #MCP #Claude #AICounsel

[Gary Harrison @Gary__Harrison]

@omarsar0 @karpathy Built it before him. Works by consensus between models for deliberation. https://t.co/Q9Hm2J9Pla

[Gary Harrison @Gary__Harrison]

@matanSF $200

[Gary Harrison @Gary__Harrison]

@arlogilbert @alexalbert__ Turn off auto compact?

[Gary Harrison @Gary__Harrison]

@donvito I like mine even better than the others

https://t.co/Q9Hm2J9hvC
True deliberation through consensus not just several models spouting nonsense.

[Gary Harrison @Gary__Harrison]

@klaudworks @alexalbert__ https://t.co/Y7N8eEkcOX

[Gary Harrison @Gary__Harrison]

@klaudworks @alexalbert__ @klaudworks you should try AI Counsel then :)

https://t.co/Q9Hm2J9hvC

[Gary Harrison @Gary__Harrison]

@klaudworks @alexalbert__ I feed in what I want or what I want fixing and the counsel will provide the answer then it can hand it off to Conductor to implement it.

[Gary Harrison @Gary__Harrison]

@klaudworks @alexalbert__ Sorry only seen this now. Usually when I’m designing something new and I want a multi model perspective that’s where the counsels steps in. Now that I’ve built Conductor, my plan is to soon join them together and have a complete autonomous end to solution

[Gary Harrison @Gary__Harrison]

@haider1 That use to be true until I built Conductor. A whole multi agent orchestration that keeps pace and has context from what’s been done and what’s next. https://t.co/Y7N8eEkKEv

[Gary Harrison @Gary__Harrison]

@haider1 All those subscriptions. Why not get the AI Counsel to help with those day to day decisions?

https://t.co/Q9Hm2J9Pla

[Gary Harrison @Gary__Harrison]

@NoahEpstein_ Happy to take onboard feedback from you guys but I think you’ll be surprised at how good it is :)

[Gary Harrison @Gary__Harrison]

@NoahEpstein_ I don’t want speed I want multi model consensus to get my idea right the first time time. I want better performing agent frameworks to get the idea refined. I want and I use AI Counsel https://t.co/Q9Hm2J9hvC

[Gary Harrison @Gary__Harrison]

@NoahEpstein_ The counsel now also have tools that allow it to read your code/files (red only) and a tree diagram of your structure is given to the counsel.
All of what I’ve mentioned save tokens for earlier decision making.

[Gary Harrison @Gary__Harrison]

@NoahEpstein_ Also because it’s linked into a local db, using semantic matching if there are similar deliberations in the db, it injects those details into the debate.

[Gary Harrison @Gary__Harrison]

@NoahEpstein_ I highly recommend you give it a try. The fact you can even use local models either through ollama, llama.cpp or LM Studio gives you the opportunity to use several AI models for decision making.

[Gary Harrison @Gary__Harrison]

@NoahEpstein_ Hey guys. About cost. No it’s not as expensive as you think. The models don’t go out of control and debate forever. They have a 3-5 (or you specify how many) rounds in conference mode or you can even do quick mode. The models come to a consensus and vote through true deliberation

[Gary Harrison @Gary__Harrison]

Try again! It should be live now.

[Gary Harrison @Gary__Harrison]

AI Counsel: Mix local (ollama/lmstudio/llamacpp) + cloud (Claude/GPT) models in one deliberation. Same model, 3 runtimes. Zero API costs. Full consensus, maximum flexibility!
https://t.co/Q9Hm2J9hvC
#AI, #MCP, #AICounsel, #Ollama, #LLMs , #claudecode

[Gary Harrison @Gary__Harrison]

@RfSharko @DeryaTR_ @karpathy No better yet not Karpathy this. https://t.co/Q9Hm2J9Pla

[Gary Harrison @Gary__Harrison]

@DeryaTR_ Better if you have the models from different genetic frameworks that offer true deliberation to work towards concensus https://t.co/Q9Hm2J9Pla
Claude, Codex, Droid, Gemini, Local Models (Ollama). Just do it.

[Gary Harrison @Gary__Harrison]

@AnthropicAI @AnthropicAI
How about using Conductor? Your codebase can be as you like. Each agent has their own context window but inter task awareness.
Success rate is around 85%

https://t.co/Y7N8eEkKEv

[Gary Harrison @Gary__Harrison]

@vasuman *agentic not genetic 😂

[Gary Harrison @Gary__Harrison]

@vasuman https://t.co/Q9Hm2J9Pla

Did it. Works well. The advantage is not only having access to different models but to the different genetic frameworks that you get with Claude Code, Codex, Droid and Gemini. It even supports local models (plans, llama.cpp, LM Studio)