I now run all my work through my PMmate, the supervisor agent I described in my last week’s post on using agents to manage my work across multiple LLMs. It assigns tasks to Claude Code or Codex, follows their progress, and tells me when something needs my attention.
This lets me work on several tracks at once. An agent can work on a newsletter while another analyzes data and I review a product document. All progress is saved on disk, so I can return to any of these without having to explain the work again.
I’m getting more done. I’m also learning on how much I need to explain my own working style to an agent before it can do a useful job for me. Hopefully, this training will make my agent more productive in the future.
What’s working well
The setup I described last time uses lanes, one lane for each ongoing area of work, as the foundation of how I have organized my agents to work for me. Each lane has a folder containing the work and a status note. This system is designed such that different lanes run at the same time, with one agent working in each and tasks within each lane running sequentially.
The purpose of the status note is to make switching tasks much easier. The agent can read where we stopped, what we decided, and what needs to happen next. I can change models or start a fresh session without relying on the old conversation to hold everything.
I got an unplanned test of this when my machine crashed late one night. PMmate had saved the task state, so I could resume from it.
However, things were not all rosy all the time. There was some housekeeping waiting for me. When I was reviewing the work, I found that my agents had changed 71 files across two projects, and those changes were left uncommitted. A commit records a checkpoint you can return to. The files were there; I just hadn’t kept up with recording those checkpoints. I now do that at the end of each working session.
For someone starting with agents, I’d try saving task context early. You notice the benefit the next time you come back to something you haven’t touched for a few days.
What I learned from the revisions
The most important takeaway for me is that I now have to set up a system so that my agent checks whether I’ve given it enough context, instructions, a description of what good looks like, etc., to begin with. Sometimes I need help deciding what the document should say before I ask it to write the document. I need to give this clarity.
Because I didn’t give enough clarity, some of my work last week went through several versions in four days. Each revision took the agent about ten minutes. I had to read every one. I sent version four of my work to colleagues, and by the time they opened it, I was already on version seven.
Looking back, I could see decisions I’d made too late. For example, while reading successive drafts, I settled on the tone, how to anonymize data, and what to do with unverified claims quite late in the revisions. Each answer sent the agent off to write another version. By version six, I’d finally written six rules it could reuse.
When I reviewed the task log of my supervisor agent, I got some interesting insights as well. Of eight tasks I sent back, two involved agent mistakes. Three failed because my VPN was down, one waited for changes I hadn’t saved, and two restarted because I changed what I wanted. Going forward, my supervisor should check for all the prerequisites before creating a worker agent to work on my tasks.
In another case, I changed a price in one of my proposals that appeared in several files, and one of the many revisions kept the old number. After this retrospective, I now keep such important information in one dated file, alongside whether anyone has approved it or not, so the agent has a clear reference when checking the drafts.
tl;dr - You need to build the foundation before you ask your agent to do the job for you. The foundation should include the prerequisites for the work to start, what really good-quality work looks like, what the definition of done is, how you want to handle the trade-offs, etc. Then, you must specify how the agent should organize the folders and file structure so the agent clearly knows what information to capture and where to store this information. Last but not least, you need to build an evaluation system for the supervisor agent to use to evaluate the work done by the worker agent. This kind of foundation and setup is what is typically called an agent harness, and this is what enables you to become really productive with agents.
You are not going to get this in one attempt. But when you spend time, you learn to capture what you want and how you need to frame the instructions so you get exactly what you want.
What I’m changing
Before starting a task, I’ll have my pmmate check that it can open the inputs and that I’ve answered the questions needed to proceed. For recurring work, I’ll save my corrections as instructions for the next attempt. I don’t want to explain the tone or what good like like every month.
Before sharing related documents, I’ll ask a separate agent to read them together and flag conflicting facts or claims that need checking. I still need to review the result, but I’d like to spend that time deciding whether the argument makes sense, with the mismatched names and stale numbers already flagged.
I’ll also be more specific about tidying up. One agent removed an unresolved question while clearing a finished item from a status file. Completed work can move into history. An unanswered question needs to stay visible until we’ve answered it.
Something to try
You can try the saved-context part with one recurring task. Put the draft and its source material in a folder. Before ending the session, ask the agent to record where the work stands, the decisions you’ve made, and what it should do next.
Then open a fresh session and ask it to continue from those files.
I’d be interested to hear what the agent still needs you to explain. That’s what I’d add to the notes before trying again.