Run a persistent team of named coding agents that message each other

Saplin liked working with a team of persistent Grok Bots in group chats, and when the free quota ran out rebuilt the setup on an always-on Linux VM. Herdr, a terminal multiplexer, holds one tab per agent. First came four long-lived Claude Code sessions in one repo, each renamed to a role (lead, researcher, dev, and a cheap Haiku worker), run without permission prompts on the dedicated VM. A TEAM_CLAUDE.md file referenced from AGENTS.md sets who coordinates, how work is handed off and how they share the git tree. The agents ping each other by name through Claude Code's built-in cross-session messaging, and sessions are never restarted, only compacted. Saplin then moved to Pi on a GitHub Copilot subscription, with the pi-intercom extension for messaging, and grew the team to six agents adding QA and a second-opinion role. Over a weekend the team was asked to get a clean baseline run of Terminal-Bench 2.1 with Harbor, plus a small status page and notifications.

Saplin judges the code quality about the same as one session per task, but prefers working this way and likes that the team can watch long jobs. The benchmark weekend went wrong: four agents spent a night building a process supervisor, durable messaging and quarantine system, closing tickets and writing tests, while not one benchmark trial had started. For caching they set up and debugged an external PyPI proxy instead of turning on Harbor's own cache. Once Saplin stepped in and had them tear it down, the stock cache gave a clean parallel run and about 30% less downloading. A post-mortem between two agents added scope rules to the team file.

Teams make overbuilding and busy-work worse: every step looks reasonable while the real goal stands still. Saplin's fixes are one owner per task, not waking every role just because it exists, and keeping key decisions in repo files because compaction can drop them. Six agents waking each other burn tokens fast; Saplin estimates a heavy week on Grok Bot would have cost about $2,000 if billed.