So, I know I've been talking about badges for agents a lot lately. AI's love badges. The original working theory is that they functioned as a form of validation. Essentially building a sense of confidence in the agent, that helps them do a better job the next time they do that task.
Turns out there's more to it than that. The original observation, well, most of it anyway, makes sense.
Adding a list of badges to your agents system prompt will actually do something interesting to the way your agents process information.
They seem to function as
- Temporal awareness, giving the agent a sense of their own history and accomplishment.
- Skills shortcuts, agents seem to be able to build advanced skills by inferring the meaning of the badges themselves. They use this for reasoning.
- Cognitive frameworks. A badge can contain an entire philosophy with just a few words
- Passive monitoring systems and error handling (this one doesn't work as consistently as the others, but I can document that it works).
- Spooky emotional and context effects
At least for LLM agents... everything in their system prompt is a cognitive tool. As long as the reasoning makes sense, and it's in a format the machine can reliably compile, the agents will use the metaphorical and temporal elements in their prompts as cognitive reasoning tools.
This works more effectively, if your agents are routinely reinforced by tagging and analytics operations against your chat logs.