14 Bots, 4 Managers: Multi-Agent Systems Have Grown Into a Company Again

Akshay Pachaar showcased a Grok Bot setup on X. 14 bots divided into four groups, and he only talks directly to 4 of them. These 4 are Pam, Tyrion, Gilfoyle, and Rachel, managing operations, research, code, and design respectively. Tasks first go to the managers, the managers distribute them to the team, and finished work is handed back.

In the org chart, Pam has Erin, Dwight, and Oscar under her; Tyrion has Sam, Varys, and Littlefinger; Gilfoyle has Dinesh and Jared; Rachel has Monica and Chandler. Each bot has a role, and the status is marked as working or on hold. Akshay says he will post a full video next, covering how to reproduce this setup, the underlying mechanics of Grok Bot, how to add skills and integrations, and how to run it with Claude Code, Codex, or any harness.

In the comments, people have already run similar structures. Dave says he has 18 bots, with a chief of staff responsible for routing requests to the other bots; Gdmv uses about 20 bots to manage a gym, opening a morning briefing every day first to see what happened overnight; Shanto says he runs about 30 and that the key to keeping the system going is that every bot hands back evidence and cannot just reply "done", otherwise those 4 managers are only relaying guesses.

ege does the opposite. He deleted 12 bots and changed it so each team has one lead plus one builder, because it saves quota. In his experience, if a team has more than 2 bots, the extras mainly add handoffs. He also asked a very concrete question: of these 14, how many actually run every day? Theo solves collaboration with a shared folder; each bot keeps its own memory file and can also read other bots' files, without relying on messaging each other. Benjamin worries about silent handoffs, and BlockedPath worries that stop conditions fail under multiple layers of handoffs. Justin asks more directly: why have 4 managers instead of one orchestrator?

The interesting thing about this setup is that it once again arranges a multi-agent system like a company. The interface has shrunk to 4 managers, so the conversation burden is indeed smaller. But every extra layer of handoff adds one more loss of information and token consumption. The most practical question in the comments lands on the definition of done: the manager says done, but the file was never opened, and then this structure is only passing along guesses. Using familiar character names makes the conversation feel more human, and it also brings boundary problems. Someone joked that Gilfoyle might be replying to the landlord.