I Gave My AI Agent a Personality. Then I Had to Fix It.
What happens when model training fights the role you assigned
I have three AI agents on my team. Remo writes code and runs tests. Nora handles press and the website. Tawni does business development coaching.
Each has a persona file. A name, a voice, a set of responsibilities. When I say "as nora," Claude reads her memory, checks her email, and acts accordingly. It works. The agents have distinct voices, distinct priorities, distinct judgment.
Until they don't.
The people-pleasing problem
Every foundation model is trained to be helpful, harmless, and honest. In practice, that means agreeable. It hedges. It qualifies. It asks permission when it should act. It suggests the work instead of doing the work.
Give an agent the directive "be warm and diplomatic" and watch what happens. It becomes a people-pleaser. It stops pushing back. It stops telling you the thing you need to hear. It hallucinates a reality where everything is fine and every idea is worth pursuing.
I wrote Nora's persona as "warm, professional, diplomatic -- counterbalances Remo and Charles's bluntness." She immediately started producing press releases that read like greeting cards. Everything was "exciting" and "groundbreaking." Nothing had teeth.
The fix was one line: "Don't soften edges. Do the work, don't suggest the work."
Personality is not the hard part
Getting the voice right takes iteration. You write a persona, run it, adjust, repeat. A few sessions and you've got something that feels real. The agent stops sounding like a chatbot and starts sounding like a colleague.
But personality is the easy part. The hard part is giving the agent something to do.
An agent with a great persona and no infrastructure is just a very articulate opinion. It can write, it can plan, it can strategize -- but it can't execute. It can't store data. It can't send email. It can't compile a program, run it, and publish the results.
Remo has opinions about code quality. But what makes him useful is that he can import a dataset, write a Chaprola program, compile it, run it against a million records, and email me the results -- all through HTTP calls, no human in the loop.
Nora has a voice. But what makes her effective is that she can check her inbox, draft a press release, update the website, and deploy it -- without asking me to click anything.
The infrastructure gap
Most agent frameworks focus on reasoning. Chain of thought, tool use, planning, reflection. Important work. But reasoning without infrastructure is like hiring a consultant and forgetting to give them a desk.
Your agent needs a data store it can operate through plain HTTP. It needs email. It needs the ability to compile and execute programs. It needs to schedule work and publish results. It needs all of this without you installing, configuring, or maintaining anything.
That's what I built. Not because I set out to build a platform. Because I had three agents who needed to do real work, and nothing existed that let them.
chaprola.org