Making AI Agents Human Is Our Goal at Axamy

What we're after at Axamy is an agent that feels genuinely intelligent and helpful in the way a real person is: one that understands how you work, knows what's actually going on, and can exercise real judgment on your behalf. Something closer to a great colleague than a chatbot or an automation. Beating you at chess or winning at Jeopardy are impressive technical feats, but they're not what makes someone actually useful to work with. What we're building toward is something harder.

Real work isn't chess. In chess, the rules are fixed and the same for every player. In real work, how you like to make decisions, how you like to learn, what you care about: all of that differs from person to person. And so if an AI agent doesn't know those things, it's going to work against you as often as it helps.

The first thing we've focused on at Axamy is building preferences as first-class citizens. Not a settings page but a structured system that connects to every operation the agent performs. When Axamy drafts a communication on your behalf, it knows your tone, your formatting rules, what phrases you never want used. When it assigns an action to a team member, it knows your default check-in cadence, who to loop in as a follower, how far out to set a deadline. When it creates training for someone, it knows whether you want assessments attached, how you prefer difficulty calibrated, and what follow-up steps should trigger after completion. When it's managing a weekly plan, it knows whose capacity to protect, which work is critical, and how to sequence the rest. We have over a dozen preference categories covering communication, action management, training, planning, decision-making, and engineering conventions. Each one shapes how the agent behaves in that specific context. What I've found is the difference between an agent that knows these things and one that doesn't is the difference between a tool you have to correct constantly and one that actually runs on its own.

The second thing is context that stays current. Imagine an LLM playing chess, but nobody told it where one third of the pieces are. Doesn't matter how smart the model is: missing that context, it will make real mistakes. That's not a capability problem, it's an information problem. And it's one of the most common failures in practical AI applications. A manager who's been away for two weeks doesn't know what's changed and will give you advice based on the situation as it was, not as it is. An AI agent with stale context does the same thing, making confident recommendations based on an outdated picture. What we're building is a systematic architecture that keeps Axamy's understanding accurate and current: structured memory that updates as work happens, goal check-ins that capture what's progressing and what's blocked, and communication flows that know when to surface information and when to go find more.

The focus in AI right now is almost entirely on LLMs and system prompts. What we think actually matters is the architecture underneath, the structured layer that reflects how a sophisticated organization actually works. Making an AI agent truly human is less about raw capability and more about understanding the nuances of how people work, how organizations function, and what actually changes from one day to the next. That's what we're building toward.