It was better than none but some good some bad, Tried to fix them,
While in the effort, found revealed way bigger problem.
After some talk,
Day 4 kicked off with high-fives and a serious sense of accomplishment. We had just compiled a video documenting everything up to Day 3. Sure, it wasn’t perfect, but it was alive! However, as I dove deeper into the project, a massive structural plot twist started to reveal itself. [cite: 1]
It was better than none but some good some bad, Tried to fix them, While in the effort, found revealed way bigger problem. After some talk, Revealed worker agents never involved. |
After reviewing the system's behavior, the harsh truth came to light: our worker agents were never actually being involved in the loop! The orchestration layer was failing. When the system finally attempted to fix the task delegation, it turned into an endless loop of approval requests. [cite: 1]
I found myself approving the exact same request three times in a row! It went a bit like this:
To top it off, Hermes didn't even seem to understand the main issue we were desperately trying to fix. [cite: 1]
And some point, Hermes lost what are we doing, we were in the middle of debugging, it asks todo list at morning. Also it somehow continue change requirements owner, I asked hermes to validate before me but certain point, hermes does not validate at all and came to me directlyAfter a lot of back and forth, I’m seeing following Asked if all done, It changed, to Human instead Hermes, Orchestration though chat did not work, had to give up. |
It became clear that orchestration through a standard chat interface was a total bust. I had to give up on that approach entirely. To bypass the communication breakdown,I manually copy and pasted the entire requirement document directly into our Kanban board to triage the tasks. [cite: 1]
The workflow was designed to be an absolute powerhouse:
Role | Responsibility
|
1. Orchestrator | Initiates the task and monitors the loop. [cite: 1] |
2. Developer Agent | Creates the actual video artifact using FFmpeg, assembly, and captions. [cite: 1] |
3. QA Agent | Validates the output (checking metadata, visual elements, and captions). [cite: 1] |
4. Hermes | Performs deep validation and restarts the loop if requirements fail. [cite: 1] |
5. Human Review | My final, golden sign-off after machine validation. [cite: 1] |
The goal was : a build-in-public journey video for "No Man Company," featuring a "popping" caption style, burned-in closed captions, and specific visual assets from Day 3 (like our retro setup summary, dashboards, and proof of work). [cite: 1]
Even with the Kanban board in place, the troubleshooting phase was there. The system ran into immediate roadblocks. Fortunately, it managed to self-correct on a few, but the same errors kept popping up like whack-a-mole. I found myself debugging constantly, despite the fact that I hadn't written a single line of code! We ran into model-related issues way too many times. But guess what? We finally found the underlying fixes. And honestly? It's still 100 times faster than human effort. [cite: 1]
Looks same error. Had to bug more times. This is yet confusing, Found fix AI helps, we yet to debug ! There were a lot fo model related issues. |
Finally, the breakthrough arrived! Every single task was marked as done, and the video was successfully created exactly as requested. 🎉 [cite: 1]
Is it ready to ship to the world? Far from it. The quality still needs immense refinement, but getting an entirely autonomous pipeline to actually deliver a finished product was the real victory today. That refinement, however, is a story for next time! 😉 [cite: 1]
