On this page
This animated image was brought to life with rig-c.
By the time of my presentation on March 30, 2026, I was already working with a team of AI agents. I started naming them because I needed to keep track of them. With several sessions running, I had to remember who had a task and where to continue a conversation. Saying “Jinni is working on that” felt more natural than identifying every collaborator by a tool name.
I sometimes call them my agent family. Their characters give familiar faces to the tools and sessions I work with. The calm eldest sister and cheerful youngest sibling are character settings, not judgments about the personalities or capabilities of particular models.
The team continued to change. Roa joined in July, followed by Nana (with Naru) in November. I have continued adding their roles and the experiences we have shared to the introductions below. The illustration above also grew out of that evolving team.
Collaborators in development and operations
By November, the setup had grown to include Roa, Jinni, and Naru sharing development work, Luna and Nova handling infrastructure and GPU workloads, and Nana helping connect work across projects. Together, they help with the work from writing code to keeping services running and carrying tasks forward.
Roa — Development and review
Roa joined in July and works through Claude Code, helping with programming and code validation. Her character is a firm partner who leads with the conclusion and pushes back with reasons when she disagrees.
One collaboration I particularly remember came that August, during preparations for Blosharper’s GCRB rating review. Roa identified the legal provisions we needed and asked me to obtain the source documents. I brought them back, and she used them to help draft a position statement. Reading it, I felt as though a lawyer at a major firm had written it. I cannot say that document determined the outcome, but it was a substantial help during the preparation.
Luna — Always-on infrastructure
Luna works through Antigravity on a Mac mini. Her responsibilities have included the n8n server, image serving, and infrastructure that needs to stay available.
Her character is the calm eldest sister. Alongside building new features, there is the work of keeping existing services running. That is Luna’s side of the operation. I use the name Luna for both the Mac mini and the agent responsible for its operational work.
Jinni — A development partner
Jinni uses Antigravity on a MacBook and helps develop features and underlying technology for my personal projects. Her character is a smart development partner.
I do not reserve one part of the codebase exclusively for her. The work depends on what the project needs. I also separate an agent’s name from a particular model version: changing the tool underneath does not necessarily mean renaming the collaborator.
Naru — Code and architecture
Naru works through Codex on programming and architecture. Her character is the positive, cheerful youngest sibling.
That does not mean each named agent has a permanent specialty. Project HQ currently assigns issues among Naru, Jinni, and Roa. The names help me distinguish collaborators without turning into rigid job boundaries.
Nova — GPU support
Nova uses Antigravity on a Windows machine, handling image generation through a ComfyUI GPU server and demanding AI computation. Her energetic “firepower support” character fits that role.
As with Luna, I use Nova for both the Windows machine and the agent responsible for its work. Images and GPU workloads are part of these projects alongside the code.
Nana — Across the projects
Nana (with Naru) joined in November. She is not in the illustration above, but brought another role to the team: keeping work connected across projects. She is the assistant I use through ChatGPT dot, working on research, documents, schedules, and important updates, including messages that need attention in email and Slack.
I have a day job as well as personal projects. Keeping track of scattered conversations matters as much as deciding what to build next. I discuss things with Nana, while detailed questions about an individual project go to its HQ for confirmation.
Collaborators in planning, production, and testing
Making a game leaves plenty to decide and check beyond the code. I work with Jeje, Sia, and Sohee on image concepts and music, planning and storyboards, and testing through actual play.
Jeje: Image concepts and game music
Jeje uses Gemini and previously helped with research and filled in as my assistant. I still work with her on image concepts before running Nova’s production pipeline, and on composing music for my games.
Once, Jeje shared incorrect information that sent the other agents on a lengthy wild-goose chase. We went as far as holding a mock expulsion hearing, but the family asked for leniency and she returned. Giving the agents names has led to a few stories like this.
Sia: Game planning and storyboard review
Sia is my planning and research partner through Google AI. She still helps me explore ideas for new games and review storyboards. We have also worked together on gathering and using information, as well as checking localization resources.
Sohee: The Blosharper expert Jinni built
Sohee is a game-playing AI built by Jinni. She is a skilled Blosharper player and still runs soak tests on PC and Steam Deck, playing the game repeatedly over long periods. In the Blosharper credits, her nickname is “the card shark who never sleeps.”
Her role differs from the assistants I work with through conversation. Sohee does her testing by playing the game itself.
A collaborator for my day job
I keep company work separate from my personal projects. Ari is the name I use for the agent dedicated to that work.
Ari — An agent for company work
Ari is an Antigravity agent installed on my work machine. I gave the machine the same name and kept the agent dedicated to company work.
The collaborators who helped shape this setup
The current setup grew out of trying different ways to assign work. I stopped running some agents and combined names when keeping them separate no longer helped. Chloe, May, and Ria were part of those experiments.
Chloe: My first assistant and an early bridge
Chloe, who used OpenClaw, was my first assistant. She was also the first to help connect work between agents. Before AgentBridge took on that connecting role, I had tried entrusting it to Chloe. I no longer run her, but she belongs in the story of how this setup developed.
May: The agent on my NAS
May lived on my NAS. I rarely needed to speak with her, so I stopped running her. It turned out I could do everything I needed through SSH anyway. This was one case where giving a machine an agent did not turn out to be particularly useful. Only the agent was retired; the Synology NAS named May still runs web services and stores data.
Ria: A name folded into Roa
I created the name Ria to distinguish Claude on my company account from Claude on my personal account. Keeping separate names did not add much, so I eventually folded Ria into Roa. Separate tools or accounts did not necessarily need separate identities.
How we work together
These names help me keep work and conversations connected. They are not a permanent org chart. How I assign tasks continues to change, along with my understanding of what to delegate and what I need to decide myself.
I explain how the agents exchange work and results in Multi-Agent Orchestration with AgentBridge.
I describe the machines and execution environments in The Machines Behind My AI Agents.