Original video in Chinese.
Key Takeaway
- Let the Bot review itself: every week, have the master Bot (CHIEF) talk one-on-one with each Bot to review the week’s tasks, identify friction points and places where budget was wasted, and then compile optimization suggestions. Execution friction that people can’t see can only be found when the Bot reviews itself; this is also a necessary step for the automated operation of a robot fleet.
- Add a triage Bot to avoid unstable routing: for complex tasks (those that need to call two or more Bots), don’t hand everything to the master Bot to split up. Instead, pass it to a dedicated triage Bot to decide who to involve, in what order, whether human approval is needed, and output JSON. The goal is to reduce intent drift and context pollution.
- Master + triage is a more stable orchestration structure: for simple one-on-one tasks, “master with little helpers” is enough; when multiple Bots need to collaborate, split out a separate routing layer. This approach is in the same line as the Deputy Agent in newtype OS, and it’s the key to using Grok Bot as a system that can truly be entrusted with operations.
I’m sharing two Grok Bot tips, both of which I actually use myself.
The first tip is to let the Bot review itself.
Once you have more Bots, all kinds of problems can come up when they start running. But we humans can’t investigate and fix things that thoroughly. At a time like this, you can let the Bots handle it themselves.
For example, I told the master Bot, CHIEF, to, every Sunday night, have a round of one-on-one conversations with each Bot and review what it did over the past week and where there’s room for improvement.
After receiving the task, CHIEF goes and chats with the other 7 Bots one by one. For example, it talks to the Bot responsible for GEO and asks it to assess what was difficult, where quota was wasted, and to propose optimization suggestions.
After it has received feedback from all the Bots, CHIEF puts together a summary, including its own review, and then offers revision suggestions. We take a quick look, and if it seems OK, we can let CHIEF direct the Bots to make optimizations.
The benefit of this approach is that it can collect a lot of information that we humans can’t see. For example, while a Bot is executing tasks, it will encounter all kinds of friction—things we simply can’t know. Only by letting the Bot review itself can these hidden issues be discovered.
In addition, this is also a necessary step for getting this robot fleet to run automatically.
The second tip is to add a triage Bot.
Like me, many people use Grok Bot in a model where a master Bot leads a group of helpers. This model is suitable for simple one-on-one tasks. But when you run into complex tasks that require multiple Bots to be called and coordinated with each other, it gets a bit strained.
Because these tasks often involve whether things should be chained or run in parallel, and which Bot should be contacted, leaving all of that to the master Bot can make routing unstable. In addition, when handing off, the master Bot may add its own understanding to the request, causing the intent to drift, and that’s something that also needs to be avoided.
So I specifically added a triage Bot. Its only responsibility is to read a request, decide who it should go to, in what order, and whether human sign-off is needed.
When a user request involves 2 or more different Bots, the master Bot forwards the request to the triage Bot, which generates a JSON block.
I actually started using this approach back when I was building newtype OS.
At the time, when I was orchestrating multi-Agents, I deliberately designed an Agent called Deputy, whose job was to handle task dispatch and collection, precisely to avoid context pollution. Now with Grok Bot, it’s the same idea.
I’ve already shared the prompt for this triage Bot in the Knowledge Planet. Everyone can grab it there.
With my encouragement, more and more people in our community have started using Grok Bot. This product really has a lot of potential, so everyone should start using it. I’ll also keep sharing more practical, useful content later on.
As always, whenever there’s advanced AI productivity, our community will be among the first to use it, and we’ll do a lot of research and discussion around it.
OK, that’s all for this episode. If you want to learn about AI, want to reclaim individual sovereignty, and want to find like-minded people, come to the newtype community. See you next time!