[Experiment 7] I built a Tycoon game to learn about Hermes AI Agents

I think weirdly, this experiments feels like it should not have been done. The reason being that it follows the same flow as the last 4 experiments:

• Provide link to docs of a specific product (Hermes in this case) into LLM to ask for advanced skill tree similar to Cyberpunk 2077
• Ask LLM to create a game around it with focus on learning experience
• Play the game lots of times to identify reasons on how to improve it
• Make changes and adjust the game
• Write up

In the end I just lost motivation. Learning through playing is kind of fun and a really great first step in understanding something, but I now feel I want to create new agents again or radically change the game experience through a 2d or 3d environment

At a technical side, the difference with this game was that the monthly limit was reached for using Fable 5 so I decided to use Cursor Grok 4.5 instead

I think to produce a skill tree, is kind of one of those holy grails of LLMs. An AI is just perfectly positioned to classify which information belong together in a document and how they are linked. So it is obvious that Grok had no problem to produce one of these skill trees

The issue is that the design and some minor functionalities just didn’t work like I wanted. So I iterated more than normal to make the game playable and in the end I used Opus 5 to make some final adjustments to it.

So from this simple experience, Opus 5 is better than Grok 4.5. Fable 5 is better than Opus 5. Fable is just so much better at understanding my request.

Also, when Opus 5 was running to make an update (on the IOS app as a Cursor AI agent), I think it completed the code and PR branch relatively quick but then it was running forever in some weird mode so I manually stopped it but it probably costed me a few dollars as I was driving so couldn’t do much. Also, it got me worried if everything would work or not but it seems so

A few new game functionalities compared to the previous games:

• Customer questions can now arrive
• The game now have sub-agents that the player need to select. It is quite a complicated part and this will probably need more in game instructions
• The skill tree now require more skill points for skills that are advanced. Eg third degree skills require 3 skill points vs less points for less advanced skills. As the total number of points are limited, this can be quite a tricky thing for the player to decide where to invest for depth

The game can be found here: https://www.buildlooplabs.com/hermes/

Example of customer question:

We can see the agents seat in the top right corner in the main game loop screen:

We can see how some capabilities requires 2 skill points (and some even more):

What I should have done differently?

I should have figured out a way to play the actual missions. It is kind of boring to just select contracts and wait (or read) for the timing to run out

I should have added a way to properly be able to compete with others. An example will be if the high score is fixed at 30 minutes of game play. Like that it is not the player that plays the longest that got the best high score.

I am now moving on to build some agents instead for the next experiment. Either on my Hetzner VPS or on my local computer

Thanks

Andreas