But my conclusion on plan mode is slightly different. I agree that plan mode itself may be a dead end, but I still believe there might be another way to achieve the same goal.

When I was building my product, I found that the biggest issue wasn't capability, but taste. The agent could build something that worked, but it often wasn't what I actually wanted. And behind that "taste" is a huge amount of implicit context — preferences, past decisions, product intuition, and trade-offs that live in my head. Distilling all of that into context takes a lot of effort, and I suspect giving it all to a single agent may eventually become overwhelming.

I've been wondering whether a better approach is to have multiple agents with different roles, prompts, and perspectives, and find a way for them to work together efficiently.

It's still just a hypothesis though. There are a lot of complicated coordination problems to figure out, and I don't have the answer yet.

It's necessary and would be good for you if you want to learn something systematically.

But for most of the normal issues, we can not rely a lot on it.

But when I try this in different ways, a common issue is that the LLM often drifts. Sometimes it just drops important issues or records more trivial info than I need. Wondering how you solve this issue in OzBrain.

The reason is I couldn't find a agent which is as good as Codex/Claude to integrate it. If there is a agent can make full use of it by routing different tasks to different LLMs by difficulty. That would be perfect.

Not sure if there is a agent can achieve this, maybe Deepseek harness have such potential.

Let yourself say goodbye to luxury enjoyment and be more sensitive to daily unexpected incidents. You will find that happiness is just close at hand.