Series
How the agent team that builds the game was set up, measured and made cheaper to run.
A08 5 min read
When the game felt like an errand chain, I asked three AI designers the same neutral question and kept every other opinion, mine included, away from them until they had answered. How that ran, and what a month of building this way taught me.
A07 6 min read
My kids' playtest notes became four pieces of work, each with its own agent, and one agent above them merging the results. Over four days that settled into a team: a root that orchestrates, owners that build, one reviewer, and a single thing that stays mine.
A05 5 min read
Three times in a fortnight I asked the agents to stop building and measure how they worked: a retrospective from the commit log, a pass through three weeks of their own transcripts, and an overnight round that timed every wait. Each time the numbers corrected something I was sure of.
A06 7 min read
In one week the agents' running costs became the bottleneck: byte limits that stalled good work, ceremony nobody read, and test suites that ran for hours. I asked where the time was going, and most of the answer was to do less.
A03 6 min read
Agents add code readily and almost never delete it. Three times in a fortnight I stopped feature work and had the agents audit what they had built, and the lesson was less about the audits than about what happened next: findings only got fixed once they became a queue of small, named jobs.
A04 6 min read
After a week of agents building fast, the repo itself became the bottleneck: too much to read and rules that lived only in prose. One morning made it easy to read; one night turned the golden rules into tests that fail.
A02 6 min read
In the cloud the game ran at under one frame a second, and a test suite that waited in real time fell apart. Stopping the clock fixed it; five days later, an honest bill for tests, reports and handoffs made the whole loop cheaper.
A01 6 min read
The first week of Brackenmere ran as several AI agents at once, each in its own worktree, each starting from one short page that said what was true. A week later the work moved to the cloud, my personal rules moved into the repo with it, and one session became five threads in a day.