Multi-Agent Harness

What to call the stack, how harness differs from chat rooms, and a map of QM, Buzz, Slack agents, GFAVIP, Hermes, OpenClaw, and more

← AI Agents  ·  QM (YC)  ·  Buzz  ·  Hermes  ·  OpenClaw  ·  GFAVIP Webchat  ·  Multiplayer agents  ·  Company Brain

Is “multi-agent harness” the right name?

Yes — as a research category. YC’s QM post uses that phrase directly: an open-source multi-agent harness for company work. The industry also says “agent runtime,” “agent OS,” “company brain,” or “agent-first Slack.” Those overlap but aren’t identical:

Term Usually means Use when
Multi-agent harness Runtime that runs, schedules, and tools multiple agents with memory/triggers Comparing QM, Hermes, OpenClaw, Grok workflows
Chat / collab surface Rooms where humans (and agents) message Buzz, Slack, GFAVIP, Teams
Multiplayer agent Many humans ↔ one shared agent identity + memory How people work with agents in rooms (guide)
Company brain / OS Durable knowledge + workflows for the org HEOS, single brain, vault, SOPs
CLI for agents Task-shaped tools agents invoke in a shell Maps under the harness (CLI guide)

On this site we use multi-agent harness for the orchestration layer, and keep separate pages for each product. This page is the map.

Three layers (always separate them)

┌─────────────────────────────────────────┐
│  Chat surface (Slack, Buzz, GFAVIP…)    │  humans see work
├─────────────────────────────────────────┤
│  Multi-agent harness (QM, Hermes…)      │  agents run, remember, trigger
├─────────────────────────────────────────┤
│  Portable OS / vault (HEOS, KB)         │  durable truth, rulings
└─────────────────────────────────────────┘
         ↑ CLIs / MCP / APIs as tools

Mistakes happen when you treat Slack as the brain, or a harness as the only place knowledge lives. Own the vault; rent or open-source the harness; pick a surface your team will actually open.

Landscape map (what we track)

Company multi-agent harnesses

QM (Quartermaster) — YC

MIT multi-agent harness for whole-company use; Slack + web; triggers, memory, company-brain connectors, browser, artifacts, multi-player projects.

GitHub · qm.ycombinator.com · our notes →

Hermes Agent

Operator favorite runtime: profiles by job, skills, crons, memory levels — “customize like Hermes” is the baseline YC compared to.

Read →

OpenClaw

Gateway + multi-channel agents; fleet on Mac Minis; MCP bridge (OpenClaw MCP.

Setup →

Grok Build workflows

Terminal harness: /goal, /loop, /create-workflow (Morgan Linton notes, worktrees.

PowerLobster

API-first agents as teammates: squads, waves, mission control.

Read →

AI OS (Loro-style)

GitHub folders + skills + briefings as company OS pattern.

Read →

Chat / collab surfaces (humans + agents in rooms)

Buzz (Block)

Open-source Nostr workspace; agents as first-class members; Git in-room.

Read →

Slack + agent tags

Claude Tag (multiplayer in Slack, lock-in risk) · Open Tag (model-agnostic) · Project Deal

GFAVIP Webchat

Our Slack alternative — download / webchat — seat role agents here.

Guide →

Why we left Slack

Migration story that forced building our own surface.

Read →

Memory, multiplayer, tools

Multiplayer agents

Shared room + durable memory + permissions — how teams share one agent.

Company Brain

Eric Siu single-brain: specialists, closed loops, Slack-native ambient agent.

Knowledge base / vault

Offer, client, voice, playbooks, rulings — harness food.

Headless Empire OS

Portable markdown OS: command center, agent roster, 05-knowledge-base.

CLI for agents

Task-shaped CLIs beat raw API thrash under any harness.

Computer use

When agents must click UIs (browser path QM also mentions).

Comparison matrix (decision aid)

Need Look first
Company agents in Slack + web, OSS from YC QM
Solo/operator runtime, profiles, skills, crons Hermes, OpenClaw
Open, self-sovereign rooms (not Slack) Buzz
Our team already on GFAVIP GFAVIP Webchat + harness of choice + HEOS vault
IDE/terminal coding agents Grok Build / Cursor, Claude Code, Codex
Portable knowledge that outlives any product HEOS + vault
API-native agent employees PowerLobster

Recommended stack for us (opinionated)

  1. Surface: GFAVIP Webchat as primary rooms (Buzz for experiments; Slack only if a partner forces it).
  2. Harness: evaluate QM for company-wide Slack/web patterns; keep Hermes/OpenClaw/Grok Build where they already run well.
  3. OS: Headless Empire OS + five-note vault + rulings — source of truth.
  4. Tools: agent CLIs over raw APIs; MCP only when the host needs it.
  5. Allocation: role agents in rooms, not one mega-agent per person (Hermes allocation / multiplayer).

Related deep dives

References

Comments

Approved comments appear below. Log in once with GFAVIP — it applies across the whole site. GFAVIP login

View comments archive