Building Atlas in the Open
Development notes from building Atlas: what we are shipping, the problems we hit along the way, and the thinking behind the decisions.
The Best Tool for AI Agents Is Eleven Years Old
git worktree shipped in 2015 and almost nobody uses it. Coding agents finally gave it a reason to exist: isolated working directories, parallel agents, and a clean per-branch review boundary. Plus an honest look at where it breaks.
Picking a Voice That Never Leaves the House
Picking a voice for Atlas took longer than building the wake word detection. Because it runs air-gapped, the real question was never which TTS sounds best. It was which one runs on my own box with no network call. Six models, one constraint, and why the best option was the fastest rejection.
The Agent That Sources Its Own Training Data
Atlas suggested giving itself a torrent client to source its own training data. I let it, inside boundaries I set. It now fine-tunes scoped versions of itself and pulls audio data on its own during heartbeat cycles. On why that is a different category of agent, and where it gets genuinely risky.
The Best Tools Show Up Before You Go Looking
Atlas scans GitHub trending every week and queues up tools that might fit its own codebase. This week it surfaced Archify, a tool I did not know I needed. On why surfacing, not answering, is the real feature, and where that quietly breaks.
16.4 Million Tokens for Six Tenths of a Cent
Atlas ran 16.4 million tokens last month, almost entirely on a local 27B model on a $1,000 box. Total inference spend: $0.0063. Local inference is not a hobby anymore, it is a deployment strategy.
Reading Is a Write Operation
I gave Atlas passive access to my WhatsApp so it could learn my conversational voice. It learned. It also spent a week telling everyone I was ignoring them.