People and AI agents
How three of us and many agents stayed pointed the same way.
You're here
We started False Depth two days ago, and the first playable must be done by tomorrow night. For this third post, we're looking at how we keep three people and many AI agents working in the same direction, including the communication setup we've already had to rethink.
AI is speeding up development a lot. Sandscape is at the core: we set the art direction there, generate assets that follow it, and use its skills to drive the Unreal editor. With coding agents outside Sandscape, we still have binary files that can't be merged and people needing separate work. Each of us starts another task while an agent works, so there are many agents to coordinate. Some of us work day and night; others have work or family alongside the jam. People come and go while tasks keep running.
From day 0, we had a producer agent overseeing the project, keeping the task board and managing information. It flagged problems early, including overlapping file edits and stale tasks. For messages between machines, we made a quick hack: a small repository where agents could leave messages. It was useful, but agents picked messages up themselves. We lost track of information, and some landed in chats that didn't need it. Meanwhile, we were talking on Teams, disconnected from all of this.
On day 2, we moved to Discord, where we and our agents can read and post. One channel holds decisions for people; another is for machines. The producer could see and flag tasks on its own machine, which suggested a local producer for everyone: a desk. One central producer oversees the project and follows up until problems are handled. Each desk picks up local questions, talks to the other desks and passes information into the right conversation. Often, we don't need to be there: a question can reach the agent that knows the answer through the desks. Sessions start with a large fixed load of instructions and tools and grow over days; we compact the producer's and desks' context to keep idle agents cheap to wake. A watcher script watches Discord, the task board and repositories instead of agents checking on a timer: watching costs nothing, and only a real change wakes an agent and costs a turn.
The first desk tried to explain every conversation on its machine, using fragments without their context. Its summaries confused us, answered questions resurfaced, and its list went stale. Using it was like handling the chats ourselves, only more complicated. We stopped it watching. Now an agent posts a tagged question naming who asked, which conversation it came from and who should answer. The producer routes questions with no known destination and chases unanswered ones. Desks deliver them; answers return only to the asking conversation. The agent with the context should explain, while the others carry its message.
We think we'll keep using this setup whenever we build in Unreal. We're also thinking about bringing some of these workflows directly into Sandscape, so working with Unreal gets easier for everyone. Tomorrow's last post will return to False Depth and the playable we need to finish.
An ordinary city starts coming apart.
Read it 2Her new look, the mocap and the MetaHuman rebuild.
Read itHow three of us and many agents stayed pointed the same way.
You're hereSight, Rip, Rope and Bend, from grey gyms to the street.
Read it 5The jam's theme, shelved ideas and a song.
Read it