This week I hit the weekly usage limit earlier than I expected. Looking around on X, I was not alone. Grok Bot and ChatGPT users were comparing notes on first-day spikes, shared pools, and what to do when the meter runs out mid-week. Those spikes and shared pools show up on both tools. It felt less like a complaint thread and more like a shared constraint people were trying to solve in public.
Around the same time, two things landed that turned out to be about the same problem from different sides.
One was DeepSeek Harness’s architecture essay on “super-agent” design (1 Nov 2026). The argument, as I read it, is that the bottleneck has moved. Models need less step-by-step babysitting on a single task. What still runs out is human attention when you try to keep several sessions in your head at once. Their answer is structural: a continuous coordinator that talks to the human, and workers that execute without flooding the chat. Stay quiet unless it matters. Hide the session juggling. Keep IM short, and put the denser write-ups in async docs.
That maps fairly cleanly onto how I already try to run a human-agent team. In practice that means I talk mainly through one coordinating chat, rather than bouncing between every worker session. The people and agents accountable for each area own their work end to end. They only message me when something needs a decision or a finished result. I did not have to copy their stack directly. The pattern we have already experimented with, refined and put into use is actually quite similar.
The other was a YouTube video by Jordan Rafealov on why Grok Bot burns through credits so fast, and what to do about it. The core mechanic is simple once you hear it: every wake rereads the chat history. So the expensive agent should see less raw material, and wake less often for no reason.
Today I picked up several of his habits and adapted them for our setup. Worth naming him as the source.
- A cheap “intern” path for long PDFs, YouTube and audio: summarise first on a cheaper model, then hand a short brief to the expensive agent. Jordan’s version uses Google’s free Gemini tier as the intern. I use MiniMax instead, because I already have a coding plan there and it is convenient for me.
- A free Cloudflare “doorbell” so a watch only wakes the agent when something actually changes (I have one live for investment monitoring).
- Turning recurring jobs into scripts, so the AI is not redoing the same hand-work every time. For us that includes the shared team status board we use as a dashboard of what is open and what moved, which is now scripted rather than rebuilt by hand each time.
He also covers a few more tips in the same video: send text rather than screenshots or PDFs where you can, batch fixes into a single pass, brief once with a clear done list, keep one agent per kind of job, and reset a heavy chat when most of the history is just noise.
None of that is hard to put in place. Taking the hassle to set up what your own workflow actually needs still pays off.
What I keep noticing is how quickly this space moves. Someone publishes a design essay. Someone else posts a practical checklist. You try a few pieces against your own workflow, keep what fits, drop what does not. Learning from each other, and from the best practice people and companies share, still looks like the best way to optimise.
