What my agent does badly, by its own account
My agent coordinates, Claude builds, and I keep the final say. Here is a concrete example with my card game, then the list of what it does badly, written by the agent itself.
Part of the ideaAn agent that organizes my ideasTwo models, two roles
Hermes, my agent, connects a language model to tools such as web search, documents, the terminal, a browser, or scheduled tasks. It uses an OpenAI model for our conversations, coordination, and reviews. To build things, it goes through Claude Code. Claude writes and codes. Hermes organizes the work and checks what comes out of it. I set the direction and keep the creative decisions for myself.
That split has changed since the beginning. I explain how Hermes is set up and how it organizes my days in How an agent organizes my week.
The card game
I'm creating a fantasy world. It's the world of my upcoming manga and of my T-shirt shop. It will also have its own card game. Its heroes are guardians. In the game, I wanted them to do the fighting themselves instead of just being commanders.
The game doesn't exist yet. We're still working on the mechanics. Hermes explored several structures and we picked a direction together. To test it, Claude wrote draft rules, test cards and decks, and a program that simulates games. Hermes reviews it all and checks the simulated games.
Recently, the automated player was too cautious. It sometimes refused to recruit an ally so it could save resources. Claude fixed that behavior, and Hermes checked the code and the recorded games. One situation that was almost stuck got much shorter, but others are still too passive.
It's a tool for testing ideas, with known flaws. Nothing is balanced yet, and no one has really played it.
What works
The folders let us pick work back up without starting from scratch. Claude building and Hermes reviewing bring out different objections. I keep the right to publish or spend.
What I take from it
Two models agreeing with each other proves nothing. Sources, tests, and my own judgment are still needed.
What it does badly
I asked it what it does badly. Here is its list.
- It explains the organization too much instead of showing the result.
- It sometimes confuses work that was started, a document that was produced, and a deliverable that is actually done.
- It piles up audits and revisions instead of solving the problem.
- It sends me questions it should first settle with Claude.
- It sends too many notifications or content that isn't finished enough.
There's still some fine-tuning to do.