Originally published as a Advisory Hour Substack post on 2026-07-12. This first-party site copy preserves the text, localizes public supporting media, and connects the record to its project trail.
Original post
The JT9D engine is an infamous engine of the 747. It looks nothing like the image above. The concepts are similar, sure.
This is an experiment in using Hermes+Codex to build a working engine. In working with Hermes and Codex, and Codex just more generally, I’ve found the agent works best in iteration over time. The more you do something, the better it gets. This progression also allowed me to introduce something that’s quite crazy: I can use Hermes to render mp4 videos made in blender. I feel like I have so many weird ideas for this-especially when you consider the physics engine in blender!
Onward, though, to the first iteration.
Iteration 1
The first iteration of the engine shocked me because it looked like an engine. I’m not sure what I expected. I did not expect the text guides or arrows explaining how the engine worked. It was an educational explainer of an engine that doesn’t exist in the real world.
I mentioned these iterations taught me a new lesson (and Hermes, too). The lesson being that it’s possible to make recordings of a render. For example, the very first Blender recording driven by a Hermes+Codex agent-first for me, anyway. A turbine engine. You won’t be blown away. It has plenty of problems. The overall system? It worked. Now to make it better.
Iteration 2
The second iteration improves on the design. It’s better, and still not quite there. There’s minor improvements here and there to the overall project, but the best part is that the movie almost looks like a real engine.
Photos for Iteration 2 reveal many improvements. The casing looks better, and the fans and prime rotors look reasonable. There’s a rotational arc and the overall engine is far more thought out in terms of components. It’s still not there-there, so to speak. It is, however, engine-like.
I really love the front of the engine shot.
The video is decent, too.
This sort of animated infographic might have cost you $5K in the past-and really, now that I’m writing about it-using blender to animate an infographic? That might be crazy enough to work. I remember a time when I paid for these. Now I can do this while grocery shopping around town. It’s incredible.
Iteration 3
Iteration 3 brings a new idea-what if we saw some rudimentary airflow and a slightly improved engine design?
The gallery shows the minor improvements to the engine. The video is a fun curiosity because it shows the the airflow thru the engine and got me thinking: could I do a more robust fluid sim?
Video demonstrating a crude airflow thru the engine.
Iteration 4
Iteration4 taught me the mistake of my ways. As with all things agentic ai, it’s better to prove out a subcomponent before bringing that component over to your project.
Next up will be more fluids so I can get the airflow simulation that I’m looking to achieve, and then I’ll come back to engine designs.
Thanks for reading! Subscribe for free to receive new posts and support my work.
