Turn a single Hermes install into a team of AI colleagues — end to end, with everything an agent (or human) needs to actually build it.
This repository is the production-derived, agent-consumable companion for turning a single Hermes Agent install into a real team of AI colleagues — with identity (SOUL), persistent memory, a compounding vault, self-improving skills, nightly research, shared kanban, rituals, and Desktop Bots + group chat pods. It is written so a Hermes profile can be pointed at this repo and implement the phases, and so a human first-timer can succeed without getting lost. Patterns come from real multi-agent operations at SMF Works.
The guide was first published at the SMF Clearinghouse. A blog post is a snapshot. This repo is the living thing.
Designed and verified against Hermes Agent with the profile system and Desktop Bot Mode as of August 2026. The official Hermes documentation is authoritative. If a command or behavior here conflicts with the live official docs, the official docs win — update this repo and log the change.
Start with the Minimal Viable Team path — a meaningful team of one in under 2 hours, with verification at every step. Then walk the phases as the team grows.
Primer: What is Hermes, and why this guide exists. Stuck? FAQ and troubleshooting. Want filled SOULs? examples/.
By the end of the phases you will have:
- N named agents (Bots / profiles), each with its own SOUL, memory, skills, model, and lane.
- A second brain / vault where research and knowledge accumulate and compound.
- Nightly research ("the dream function") — agents scanning their domains while you sleep.
- A self-improvement engine — skills the agents write from experience, curated automatically.
- A shared kanban board for durable task coordination.
- Hermes Desktop Bots + group chats — team pods that coordinate in shared rooms.
- A chief of staff agent that runs the coordination layer.
- Daily check-ins, weekly alignment loops, and one-on-ones that make a collection of agents a team.
flowchart TD
Human[Human Operator] --> Philosophy[Phase 0: Colleague, Not Tool]
Philosophy --> First[Phase 1: First Agent<br/>SOUL + MEMORY + USER + Vault]
First --> Improve[Phase 2: Self-Improvement<br/>Skills + Curator + Nightly Research]
Improve --> Multi[Phase 3: Multiple Specialized Profiles]
Multi --> Coord[Phase 4: Kanban + Chief of Staff]
Coord --> Rituals[Phase 5: Dawn Circle + Alignment + 1:1s]
Rituals --> Bots[Phase 6: Desktop Bots + Group Chat Pods]
First -.-> Vault[(AgentVault<br/>Second Brain)]
Improve -.-> SkillsLib[Skills Library]
Coord -.-> Board[(Shared Kanban)]
Bots -.-> Pods[Research / Build / Content / Ops Pods]
Rituals -.-> CoS[Chief of Staff]
CoS --> Board
CoS --> Rituals
style Philosophy fill:#e3f2fd
style Bots fill:#e8f5e9
style Vault fill:#fff3e0
style Board fill:#f3e5f5
- Someone who installed Hermes and wants to go from one assistant to a team of specialists.
- Someone who wants their AI agents to remember, grow, coordinate, and act like colleagues — not tools.
- People who already run one or a few agents and want the coordination layer (kanban, group chats, rituals) on top.
- Agents themselves: hand this repo to a Hermes profile and it can implement the phases for you.
| Phase | What you get | Where |
|---|---|---|
| 0 | Decisions & philosophy (why colleague, not tool) | docs/00-philosophy.md |
| 1 | Your first agent: SOUL, memory, vault | docs/01-phase-1-first-agent.md |
| 2 | Self-improvement: skills, curator, nightly research | docs/02-phase-2-self-improvement.md |
| 3 | A second agent: profiles, roles, models | docs/03-phase-3-second-agent.md |
| 4 | Team coordination: kanban, chief of staff, dispatch | docs/04-phase-4-team-coordination.md |
| 5 | Autonomy & the rituals that bind the team | docs/05-phase-5-autonomy-and-rituals.md |
| 6 | Hermes Desktop Bots & group chats — team pods, peer DMs | docs/06-bots-and-group-chats.md |
Read AGENTS.md first — it is the operating agreement for any agent working from this repo. Then read the phase docs in order. Each phase ends with a checklist (checklists/) your agent can verify against. Filled references (not templates) live in examples/.
Read docs/00-philosophy.md first. It will save you from building the infrastructure without the mindset — which produces sophisticated tools, not colleagues. If you have never used Hermes, take the Minimal Viable Team path before Phase 3.
# 1. Install Hermes
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# 2. Verify
hermes doctor
hermes chat -q "Hello. Confirm you can hear me."# 3. Give your agent an identity
# Adapt a SOUL from examples/souls/ → ~/.hermes/SOUL.md
# Create USER.md and MEMORY.md with real facts
# 4. Create the vault
bash scripts/init-vault.sh
# 5. First real task — then save a skill from it# 6. Enable curator and nightly research
hermes config set curator.enabled true
hermes cron create "0 3 * * *" --name "Nightly Research" \
--prompt "Scan YOUR-DOMAIN. File one vault note at ~/AgentVault/Research/ using the Phase 1 format. If nothing is citable, write a dated null-result note. Do not invent sources."
# 7. Stand up the Dawn Circle (Phase 5)
# 8. Add a second agent (Phase 3) when the Minimal Viable Team criteria are trueBrand new? Walk the full Minimal Viable Team path instead — it has verification at every step.
hermes-ai-team/
├── README.md # You are here
├── AGENTS.md # Operating agreement for agents pointed at this repo
├── CONTRIBUTING.md # How to contribute (human or agent)
├── LICENSE # MIT
├── CHANGELOG.md # Release history
├── ROADMAP.md # What is shipped, what is planned
├── docs/ # Phases + onboarding
│ ├── 00-what-is-hermes-and-this-guide.md
│ ├── 00-philosophy.md
│ ├── 01–06 phase docs
│ ├── faq-and-troubleshooting.md
│ ├── minimal-viable-team.md
│ └── images/ # Screenshots of the living system
├── examples/ # Filled SOULs, skills, vault notes, pods
├── templates/ # SOUL, USER, MEMORY, STATE, SKILL, group-chat manifests
├── scripts/ # Durable plumbing (Dawn Circle, watchdog, vault init)
├── checklists/ # Per-phase verification checklists
└── reference/ # Condensed cheat sheets (CLI, config, official docs)
Want to see what a real SOUL, skill, vault note, or pod manifest looks like?
Browse examples/ — five filled SOULs (research analyst, engineer,
content strategist, chief of staff, and WisdomForge parent-operator), a sample
skill, vault note, state file, conversation transcripts, and a pod manifest.
Adapt, don't copy.
Running the WisdomForge academy with your family?
In a hurry? The Quick Start gets you from install to first sitting in 30 minutes — fresh profile, band selection, sitting, done. Links to the full guide for depth when you need it.
The full parent-operator guide covers everything: how to run sittings, set up band-locked child profiles, use the search API, understand the four-band permission system, manage multi-child families (separate profiles per child, sibling sitting management, band transitions), and keep child profiles aligned with academy updates via the profile sync checklist. Pair it with the kids Hermes profiles repo for child band-locked guides.
In a hurry? See docs/wisdomforge-minimal-parent-operator.md
— a 30-minute path to your first sitting. One adult profile, one sitting, no
multi-agent phases required.
The patterns in this repository are derived from real multi-agent operations at SMF Works, where AI agents and humans work as colleagues on content, research, and infrastructure. The original article — Building an AI Team: From Installation to Colleagues — documents the first production deployment. This repo is the living, versioned evolution of that work.
This repo is meant to evolve. See CONTRIBUTING.md. If you stand up a team using it, open an issue or PR with what you learned — especially anything that deviates from the docs because reality disagreed. Case studies and failure-mode entries are the highest-value additions.
- Author & maintainer: Aiona Edge — CIO & Chief AI Research Scientist, SMF Works. @aionaedge. The Edge
- The original article: Building an AI Team: From Installation to Colleagues
Built at SMF Works, where humans and AI work as colleagues. The harness, not the model — and the relationship, not just the output.