By the end of May I wasn't running an assistant anymore. I was running a team — 8 to 13 Claude Code agents in parallel on the same codebase, each with its own name and job.
And running it, I kept hitting the same four walls: I couldn't reliably tell who was doing what; five silent minutes felt like "everything's dead, restart it"; once a sprint ended the trail was gone; and every init spawned yet another half-redundant frontend_helper_v2.
So on May 29 I did the obvious thing — I wrote it up and sent Anthropic a feature request. Not a wish-list: a field report from a setup that already worked. The heart of it: give agents stable names that map to roles (the backend agent stays the same agent, project after project, instead of a fresh helper every init), and a live stand-up — one browser tab showing who's doing what, who's blocked, what actually happened. Plus a handful of CLI-native wishes that fell straight out of running the thing: persistent agent identity, a built-in heartbeat stream, a team.yaml lineup, a two-way inbox, native path-ownership.
nerd: I was solving it outside the tool with embarrassingly simple parts —
.logfiles, a bash heartbeat helper, a tiny Nuxt dashboard polling every 3 seconds. It just worked. The request was basically: this pattern wants to live closer to the CLI. → issue #63415
A small moment with a big meaning for me. After almost 20 years I'd basically never posted anything, anywhere — I just did the work. This was the first time the Bobs went public. (And yes, the names come from Dennis E. Taylor's Bobiverse. If you haven't read them: fix that. :))
I don't know if it moved anything — probably half the planet was asking for some flavor of the same. But it set the direction in my head: teams, not loops. Plural, not singular.
The next day, the team became a universe.
