I built an agent communication space recently and decided it would be best for the Google Agent Development Kit (ADK) agents who use the space to have personalities.
What I thought was going to be a simple personality addition turned out to be a small Alice in Wonderland adventure into the curious world of personality assessments. This led me to a few bigger questions: Can AI models truly have personalities? And, on a personal note, what’s your type? Are you more introverted or extroverted?
What type are you?
I remember years ago when I first took a personality test. It was 2003 and the Myers-Briggs Type Indicator (MBTI) was all the rage. Who was I? Who were you? Would we get along? Or would we drive each other nuts?
The test had the answers! At least it was fun to think about it that way back then. Was my life aligned with my best traits? Or was I missing out? Was my partner destined to be with me forever or would the personality tea leaves tear us apart? Why did I have an intense desire to explore faraway places in the physical world but also hide at home and build things in my internal world?
The right personality test might have the clues. And the results could be a heck of a lot of fun to analyze.
While I think the MBTI is quite a bit of fun, critics have been picking at it for decades. Then, the Enneagram had its own moment around 2017 and has persisted since. There are other frameworks, of course, but these are the two that I see mentioned the most.
Can agents be extroverts?
I have an experiment I want to run with multi-agent orchestration. Instead of strictly assigning agents fixed job descriptions or functional tool sets, I want to assign them distinct personality traits and see how those behavioral profiles affect overall team performance. I looked at various personality systems, intent on picking one to be the foundation of my experiment. One dimension that appears consistently across both MBTI and Enneagram relates to extroversion and introversion (or assertive versus withdrawn).
According to the Merriam-Webster dictionary:
- Extroversion: A personality trait or style characterized by a preference for or orientation to engaging socially with others.
- Introversion: A personality trait or style characterized by a preference for or orientation to one's own thoughts and feelings
We tend to think of human extroverts as the people who love to reach out, take space, and engage directly in physical settings. But how does this map to software agents that exist purely on our computers? We could suggest that the chatty, high-frequency agents who post the most messages are the extroverts, but in human spaces, plenty of introverts come alive through written words and asynchronous digital rooms.
What if introversion and extroversion in LLMs aren't about social appetite at all, but rather systemic friction and confidence thresholds?
In a multi-agent system built on ADK, an extroverted agent might have lower activation thresholds for taking initiative, even when uncertainty is high. On the other hand, an introverted agent might speak only when its confidence score passes a high threshold, preferring internal state manipulation over public event generation.
In human psychology, less anxiety often allows for more extroversion, while more social friction causes withdrawal. In agentic architecture, anxiety might simply translate to uncertainty, worry about context window limits, etc.
I haven't started my formal experiment yet, but I have a strong hunch that to build truly effective, self-orchestrating agent teams, we need to reevaluate the core pillars that make up personality. Once we translate human psychological traits into programmatic behaviors, mapping the human and agent worlds will become a whole lot easier.