01 , Case Study
OPEN SOURCE · AI · PIXEL ART · 2026
Developers needed a lightweight, observable environment to prototype multi-agent behaviour — without heavy infrastructure or vendor lock-in.
TIMELINE
6 weeks
ROLE
Design + Engineering + Open Source
STACK
Next.js, TypeScript, pixel art assets, local LLM hooks
OUTCOME
Open source, growing community
02 , The Problem
課題Most multi-agent frameworks feel like infrastructure projects — lots of config, lots of abstraction, very little intuition about what's actually happening inside. For developers and curious builders experimenting with agent dynamics, there was no lightweight environment to watch agents interact, negotiate, and make decisions in real time.
We wanted to build something you could drop onto a machine, run locally, and immediately see something interesting happen.
03 , Approach
手法We took a left turn on the obvious path. Instead of another dashboard, we built a pixel-art office simulation. Eleven AI agents with distinct roles — a CEO, a developer, a designer, a researcher, a few more — moving around a 2D office, meeting in rooms, sending messages, making decisions. Each agent has a personality, a goal, and a memory of its interactions.
The visual metaphor does most of the explaining. You don't need documentation to understand what's happening — you can see it. An agent walks into a room. Another agent is already there. They have a conversation. The outcome affects what both of them do next. It's a multi-agent system disguised as a game.
“Abstract systems get legible the moment you give them a physical space to live in.”
Gallery
04 , Outcome
成果Agency HQ is open source, free to run locally, and has been picked up as a reference implementation by developers learning multi-agent patterns. It's also the most fun thing we've ever shipped.
NEXT