Main Agent does not follow explicit instructions
My profession is a manufacturing engineer. I specialize in the CNC trade. The only programming I do is in this trade. I’m no a C++ programmer or anything close to it. I’ve been messing with computers since the 90’s. It’s more of a hobby than anything. But I have decent understanding of how a lot of things work.
When CNC programs go to the machine for the first time, the machinist will prove them out. When they successfully make a part and are happy with the outcome they will upload the program back tot he network for my review. I usually compare the original with the proven program, review their changes and determine what should and shouldn’t stay. Once it passes my inspection I will remove a safety header in the program telling the operate program has not been proven, change the program name by changing the files ext, update the tool list in the program if tools were moved around at the machine for various reasons, then sign, date while listing any changes at the bottom.
I’m currently learning openclaw and how to use it. I’ve set up a simple server. I have openclaw on a vps connecting to my rig at home through tailscale using Ollama. I have three agents. My main and two sub agents. I want to set up a simple workflow to do the work i mentioned above. I send the two files to my main agent with the phrase “review proven programs”. This kicks off a workflow. The main agent is to drop the two files in the first sub agent I temporally named local-coder. That agent will do the work i stated above crates a new program with the added word “approved” in the file name along with a txt file listing all the differences between the two programs. That agent passes the files to the 2nd sub agent to review all work done and flag any mistakes or errors. For now, I have a loop of 3 times max before it stops and the main agent reports back to me failure. I then will review the files myself. Regardless, i always review these files after completion.
My issue is, I’ve explicitly told my main agent time and time again that it’s only the traffic cop. It only handles the files from me, to the agents then back to me. It is not to ever review the files itself. Im trying to make it so the sub agents are the ones doing their role for consistency. unfortunately, the main agent continues to review the files itself whenever something doesn’t go right. It simply just doesnt remember. I have it in three different .md files but it doesnt matter. It seems like it’s even worse if i start a new session.
What am I doing wrong? I would think this kind of workflow is sinple and local models can handle it. Main agent uses qwen3.6:27b, coder uses qwen3-coder and reviewer uses qwen3:14b.
Appreciate who ever took the time to read this.