Here is what I'm trying this month -- April 2026.

After last month, I was tired of spending soo much time reviewing and editing docs. It didn't feel like I was accomplishing a whole lot, and I was still stuck making iterative changes after it was done.

This month, I've been trying something a bit different. For most projects, I've focused on using two agents.

In the main agent thread, I start by discussing my idea. I give it instructions that it's primary purpose is to help me brainstorm and idea, with an expected output of a file. I often leave it a little open ended. Interestingly enough, it often ends up creating a numbered design file. This is likely due to context added in previous month attempts.

During this thread, I talk about my idea, ask for ideas, and have an open discussion until I'm ready to start building something. I try to focus on small projects, so this hasn't been a long discussion. Once I have a file, I'll move to the second agent.

The second agent is a fresh start in another panel. Within this agent, I have it implement the given feature. I'll typically work within this agent chat until I'm happy with the implementation. This often includes me playing with whatever is built, giving it errors or changes that I want to make. Once I'm generally happy, or have an idea for a major revision, I'll exit that chat.

Then I move back to the main chat thread. I'll tell it that I have completed that step, and then give some other direction. Sometimes I ask to review what was done, sometimes it's to discuss the next thing that I want to do.

I've found a few interesting things while I've been doing this that seem like good lessons learned.

First is to have the main agent create a todo.txt for what it wants to do. I usually have a design too, and have it reference it back and forth. This gives me strict steps for the implementation chat. I have that chat complete the tasks and remove them. I have found that this can allow me to move back to the main chat while a sub chat is performing it's implementation.

Second is to have the main agent create the starting prompt for the developer. I say something along the lines of "create me a prompt to pass to the implementation agent to perform this work". This seems to work better than me firing up an agent and asking it to implement said features.

Finally is to exit the implementation thread at some point. I think this helps keep it from getting stuck in poor implementation loops, which I have seen happen a few time with longer implenetation threads. This doesn't seem to be as big of a problem within the ideation chat -- I expect it's because that context is generally smaller than one that is potentially spreading itself out across code repos.