@Keith Motte Thanks Keith, great insights. I saw that OpenAI is much friendlier to OpenClaw since they acquired Pete. On my new OC setup I connected the gateway to my openai account via OAuth so we'll see how that goes. It's the $30/mo business tier seat. I'm optimistic that the default model will be gpt 5.4, (which is too costly to run default via API) but I'm not sure how/when any limit caps will kick in. I did try offloading some tasks to Qwen and lighter LLM models with my first build. One hurdle I ran into was that my main agent's heartbeat directed him to check for tasks in the PM tool and then if there was work to do, to start on it. Doing this on a lightweight model meant that the task work it did during heartbeats was lower quality. The solution was to change the heartbeat to only CHECK for tasks, and then when work was discovered, pass it off to a subagent or other specialized agent. Working on forcing new sessions to keep the context window fresh/clean as often as I can to keep the cache context clear.

@Carlos Jimenez I plugged the new openclaw setup into my chatgpt oauth and it's been great. I do have to watch my weekly and 5hr usage limits, but even if I have to switch to API after I exceed the limits, that's already a huge cost savings. I have not tried Minimax, but I'll give it a shot. Have you tried Qwen, Kimi, or Nemo local models?

I have Qwen on my mac mini and it's pretty good. I have my agent route any simple or routine task to Qwen, either the 7B or the larger (slower) 32B model. So far I've had it do basic scraping, system checks, stuff like that. I plan to experiment with it more to see how much more I can do locally vs from the bigger smarter LLMs

@Keith Motte Yeah I am pretty intrigued by the new Nvidia one too.

@Keith Motte Great insights, thanks. I set up the new one last night so today I'm going to draft new governance and start fresh loading skills and building mission control. I'm going to try Obsidian for the detailed logs and interconnected wiki memory. Any thoughts or resources to point me to regarding : 1. Agents building agents or skills? An agent programmed to use best practices when birthing new specialist agents and workflows. Or best to draft in separate LLM (what I've mostly been doing lately). 2. Skills vs project workflows. And how/when these are pulled into context so they are present when the guidance is needed but not bloating context unnecessarily. Some things don't need to be boot loaded in agents md unless the request calls for it. (I hesitate to let the agent decide bc my last one kept pulling in all this expensive context bloat but hate to micromanage a system that is supposed to know how to set itself up.) 3. Memory. The big mystery. Best way for main OC agent to save memory and load it into context when relevant without doing so unnecessarily. Sorry if this should be a new topic thread instead of here in this one :)

@Jules B I do like the idea of 2 separate environments and the option for them to assist one another in some way