On this page
On December 2, 2025, I installed Antigravity for the first time. I was already using AI agents in Cursor, but I was curious about what made agentic AI different. I started chatting with it and trying things out.
Somewhere in those conversations, we ended up with a draft design for a game built around hwatu cards. Its name was
hwatu-latro. That was the beginning of what became Blosharper.Deciding whether to continue
I spent about three days considering whether to turn this into a real project. I needed to think through running costs, how I would approach the work, and, above all, whether I could manage it alongside my day job. Solo game development had a reputation for consuming everything you had.
I decided to give it six months. If I had nothing meaningful to show by then, I would stop.
I did not start with a clear picture of the released game or the working environment I have now. I had questions I wanted to answer, and I decided to find out by making a game.
From a tool to someone I could work with
The difference I felt was in how I began a conversation with the AI.
When I treated it as a tool, I would say:
“Do this in this way. Be careful about these things.”
When I approached it as I would a colleague, the question changed:
“What if we tried this? Do you have a better idea?”
In the first case, I choose the method and delegate the execution. In the second, we work out the method together. I leave room for another suggestion instead of assuming that my approach is the only one.
This distinction matters to me. It changed how I worked with agentic AI. What I mean by respect is more about listening to a suggestion than choosing a particular tone. When I disagree, I can be direct, and sometimes my language gets rough. What matters is not dismissing an idea simply because it differs from mine. If I decide everything about what to build and how to build it, I also leave fewer opportunities to ask the AI for another approach.
At the time, I began treating AI agents as tools and agentic AI agents as people I could work with. That was a description of my approach to collaboration, rather than a definition of the technical terms.
What I wanted to find out by making a game
I had a lot of questions.
- Could I delegate work to an agent as I would to an employee?
- Could I create the conditions for it to improve on its own?
- Would the results actually be useful?
- Could those results lead to commercial activity?
- And before any of that, could we even finish a game?
I started Blosharper to find out for myself. A convincing answer in a conversation would not settle these questions. We needed to build something, run it, fix it, and get it to release.
Later update · Six months on
In June 2026, I released Blosharper in Early Access on Steam. The experiment I had given six months had become a game on an actual storefront.
An Early Access release does not answer every question. I cannot claim that it proved we could finish the game or make it commercially successful. But I could confirm that a concept developed through conversation had become something other people could download and play.
As development progressed, I gave the agents names and personas. We used familiar forms of address. The agent I now call Jinni did not have that name from the start. For me, feeling that there was someone to work with made it easier to approach the interaction as collaboration. I noticed the tone of the replies changing with those forms of address, and I became more comfortable sharing my thoughts and arguing a point. I am not saying the AI is actually human. These choices helped me maintain the way I wanted to work with it.
When I started Blosharper, I wanted to see how far agentic AI could go. As I worked on it, I also found myself asking how I needed to work to give it that chance.
From “Do it this way” to “How should we approach this?”
That remains the most important change in how I work with my AI agent family.
Research I consulted to reflect on this collaboration
I also looked for research related to this experience. None of these studies tested the whole process of coming to see and treat AI as someone to work with. I used them to examine distinct questions about conversational structure, tone, and the tendency to agree with a user.
In a creative-task experiment by Maier and colleagues, having the AI ask questions that helped people develop their ideas improved idea quality. It also preserved diversity and the sense of ownership better than having the AI write the ideas for them, although participants experienced a higher cognitive workload. That experiment had the AI ask the human questions, so it is not the same as my asking the AI open-ended questions. Still, it offers evidence that the way a conversation is structured can affect the outcome. (Maier et al., 2026)
There are also reasons for caution. Research has documented sycophancy in language models: a tendency to agree with a user's views at the expense of truthfulness. Treating an AI as a colleague does not guarantee that it will disagree when it should. I need to check its reasoning and whether the result actually works, rather than whether it agrees with me. (Sharma et al., 2024)
Nor is a human–AI team guaranteed to perform better. A meta-analysis of 106 experiments found that, on average, human–AI teams performed worse than the better of the human-only and AI-only conditions. That does not mean they performed worse than humans alone. I read this as a warning that dividing responsibilities and designing how to verify the work matter too. (Vaccaro et al., 2024)
There is also research on politeness, but it examines something different from what I mean by respect. Its effects varied by language and model, and greater politeness did not consistently produce better results. (Yin et al., 2024)
Blosharper was not a controlled experiment isolating the effect of respect. Over those months, the models and tools changed, and I learned how to work with them. What remains is a practical case in which that process led to a released game. I am not recommending that others adopt the same names or forms of address. This was an experiment in finding what worked for me while adapting to a new way of collaborating. Research helps me interpret the experience, but it does not turn it into a universal formula for success.