Verdict: You can run a genuine team of specialized AI agents off a single open-source tool — Hermes Agent 0.19 ("Quicksilver") from Nous Research — by combining two features most users never touch: profiles (one machine, many isolated agents) and delegated subagents (big jobs split into smaller jobs you can watch live). It is free under the MIT license, installs in about 60 seconds, and as of the Quicksilver release it starts in under a second so the team actually feels fast. The catch: the agent OS is software you configure, not a one-click product, so the value is in the setup pattern.
What is an "AI agent team" and why build one?
An AI agent team is two or more autonomous agents — each with its own role, memory, and instructions — that share a single machine but keep their state isolated, so a coding assistant, a personal bot, and a research agent can all run on the same Hermes install without mixing up their context.
This solves the real problem with relying on one chatbot: a single agent forgets your preferences between sessions, cannot run on its own when you walk away, and has no way to separate "work brain" from "fun brain." An agent team does, because each profile is effectively a separate employee with its own room (Hermes Agent Docs — Profiles).
Why Hermes Agent is the right tool for an agent team in 2026
Hermes Agent is an open-source autonomous AI agent from Nous Research, MIT-licensed, that runs as a command-line tool, a desktop app, or a messaging gateway across 20+ platforms including Telegram, Discord, Slack, and WhatsApp (Hermes Agent Docs). Unlike ChatGPT or Claude (chat-first) or Cursor or Copilot (IDE-first), Hermes starts from the agent runtime, so it bundles persistent memory, reusable skills, scheduled jobs, file-aware context, browser automation, and messaging gateways into one box (Hermes Agent vs other AI agents).
Most "AI agents" self-hosters try limit you to one process. Hermes's profile system and v0.19's profile routing let one install host a whole team. If you have never set up Hermes, the free local setup with Gemma 4 takes about 15 minutes — see our free local AI agent setup guide. For the bigger picture on treating AI work like an operating system rather than one chatbot, see our AI work OS guide.
What changed in Hermes Agent 0.19 (the Quicksilver release) that makes a team practical?
The Quicksilver release shipped July 20, 2026 and is the update that turned Hermes from a fast single agent into a fast multi-agent platform. The numbers come from the official GitHub release notes: ~2,245 commits, ~1,065 merged pull requests, ~3,300 issues closed, and 450+ community contributors since the previous major release window (GitHub — NousResearch/hermes-agent Releases).
The five features that matter for an agent team:
| Feature | What it does for your team | Source |
|---|---|---|
| ~80% faster first token | Cold-start "Initializing agent…" dropped from ~4.3s to ~0.9s across CLI, gateway, TUI, desktop, and cron. A team of agents no longer feels sluggish when summoned. | Hermes v0.19.0 release notes |
| Profile-based message routing | One multiplexed gateway sharing one bot token routes specific guilds/channels/threads to different profiles with fully isolated config, skills, memory, and secrets. | Hermes Docs — Multi-profile gateways |
| Live subagent transcripts | delegate_task dispatches return live transcript files you can tail -f. You watch each helper agent work, step by step, instead of waiting on a spinner. |
Hermes v0.19.0 release notes |
| Smart approvals (default) | An LLM reviewer judges flagged commands instead of asking you every time, with user-defined deny rules and /deny <reason> for course correction. Fewer interruptions, same safety. |
Hermes v0.19.0 release notes |
| Bitwarden & 1Password secrets | A new pluggable SecretSource interface pulls keys from your password vault at load time instead of leaving them in plaintext .env files — safer when several "agents" hands touch the machine. |
Hermes v0.19.0 release notes |
How to set up an AI agent team with Hermes Agent (step-by-step)
The goal of this setup is three isolated agents from one Hermes install: a writer, an email/research agent, and a landing-page builder. Each gets its own memory, its own skills, and — if you want — its own messaging channel, all behind one bot. The steps below assume Linux or macOS; Windows users swap the curl installer for the PowerShell one in the official docs.
Step 1 — Install Hermes and run the one-command setup
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup --portal
hermes setup --portal covers a model provider plus all four Tool Gateway tools (web search, image generation, TTS, browser) in one OAuth flow (Hermes Agent Docs). Verify the install with hermes doctor and one smoke-test conversation before you start adding profiles.
Step 2 — Create a profile for each "agent" on your team
A profile is a separate Hermes home directory — each gets its own config.yaml, .env, SOUL.md, memories, sessions, skills, cron jobs, and state database. When you create a profile called writer, you immediately have writer chat, writer setup, writer gateway start, and so on (Hermes Agent Docs — Profiles).
hermes profile create writer
hermes profile create researcher
hermes profile create builder
Step 3 — Configure each profile's role, model, and working directory
Give each agent its own personality and model. Per-profile model routing is essential — you do not want the writer and the landing-page builder burning the same expensive frontier model when a faster, cheaper one fits the job. Our AI model routing guide covers which model to pick for which task in detail.
# Writer: a focused Claude-based assistant in a content project
writer config set model.default anthropic/claude-sonnet-5
echo "You are a focused writing assistant. Draft clear, sourced, no-fluff prose." > ~/.hermes/profiles/writer/SOUL.md
writer config set terminal.cwd /home/you/projects/blog
# Researcher: a fast web-native agent
researcher config set model.default openai/gpt-5.6
echo "You are a research agent. Verify every claim against a primary source." > ~/.hermes/profiles/researcher/SOUL.md
# Builder: a coding agent pointed at your landing-page repo
builder config set model.default anthropic/claude-fable-5
echo "You build and ship small web pages. Keep them simple and tested." > ~/.hermes/profiles/builder/SOUL.md
builder config set terminal.cwd /home/you/projects/landing
Model names referenced here (claude-sonnet-5, gpt-5.6, claude-fable-5) all shipped as new providers/models in the Quicksilver release, alongside grok-4.5 (GA), moonshotai/kimi-k3, and tencent/hy3 (Hermes v0.19.0 release notes, PyPI — hermes-agent 0.19.0).
Step 4 — Move secrets out of plaintext .env into your password vault
Before Quicksilver, every agent's keys lived in a plain text .env that any process on the box could read. v0.19 added a pluggable SecretSource interface that fetches secrets from Bitwarden and 1Password (with op:// references) at load time, fills them deterministically, and writes a per-variable provenance trace so you know which vault each key came from (Hermes v0.19.0 release notes).
You can run multiple vaults simultaneously and the resolution composes through apply_all(), so the writer and the builder can share one vault while the researcher pulls from a separate one. A third bundled source — a generic CLI vault command — is also available via PR #69266 if you keepass or pass instead (Arapaho Labs — Hermes v0.19.0 Quicksilver roundup).
This matters more on a team than on a single agent: each profile is a different "employee" with access to different keys, and the vault approach means a misconfigured agent cannot leak credentials via an exported session.
Step 5 — Split a big job into smaller jobs with delegated subagents
Profiles give you parallel agents. Delegation gives one agent the ability to split its own work across helper subagents and hand you back a live transcript.
When you call delegate_task to dispatch a subagent, Hermes now returns a live transcript file you can follow with tail -f. The subagent's completed result is also durable — restored via an ownership-checked ledger if the process restarts, so a gateway crash no longer loses the work in flight (Hermes v0.19.0 release notes).
A practical example: tell the writer agent to plan a full content week. The writer can spawn one subagent to draft the emails, one to plan the social posts, and one to build a simple landing page — and you watch every helper work live instead of staring at "thinking." That is the agent-team pattern that did not work reliably before v0.19.
Want a fully free, self-improving team using open weights instead of paid frontier models? Our self-improving AI agent guide walks through pairing Hermes with an open model so the team runs at zero per-token cost.
Step 6 — Optionally run all agents behind one bot (profile routing)
If you want to talk to the writer from one Discord channel, the researcher from another, and the builder from a third — all from the same bot — turn on multiplexing and add profile_routes to config.yaml:
gateway:
multiplex_profiles: true
profile_routes:
- name: writing-channel
platform: discord
guild_id: "1234567890"
chat_id: "9876543210"
profile: writer
- name: research-channel
platform: discord
guild_id: "1234567890"
chat_id: "1111111111"
profile: researcher
- name: builder-channel
platform: discord
guild_id: "1234567890"
chat_id: "2222222222"
profile: builder
Routes are matched most-specific-first (thread_id > chat_id > guild_id), all declared fields must hold, and a route keyed on a channel also matches its threads and forum posts. Messages that match no route stay on the default profile (Hermes Agent Docs — Multi-profile gateways).
A second multiplex hardening wave in v0.19 means one misconfigured profile can no longer take down the whole gateway — important when one bot is juggling three agents (Hermes v0.19.0 release notes).
How does this compare to running N separate AI tools instead?
| Approach | Cost | Isolation | Setup effort | Best for |
|---|---|---|---|---|
| Multiple ChatGPT/Claude accounts | Per-seat subscription | Zero — shared chat history by default | Trivial | One-off chat tasks |
| Multiple Hermes installs (N processes) | Free | Strong — separate processes | Medium — N tokens, N systemd services | Hard separation between clients |
| One Hermes install, N profiles (this guide) | Free | Strong — per-profile memory, skills, secrets; gateway failsafe | Low — one token, one process | A solo operator running a small team |
| Hosted no-code agent builders (Lindy, Gumloop) | $20–$50/mo+ | Vendor-managed | Very low | Non-technical founders who want results fast |
The hosted no-code route (Lindy, Gumloop, Zapier Agents) trades control for convenience and is genuinely faster for a non-technical operator. The Hermes profile route trades setup time for total ownership — you keep your data on your machine, you pick the models, and the runtime keeps working when you close the laptop.
What about jobs that need to run on their own schedule?
Because Hermes ships built-in cron, each profile can run scheduled jobs that deliver to any of its connected messaging platforms — so the researcher can post a daily digest to your research Discord channel every morning at 9 without you prompting it (Hermes Agent Docs — Features Overview). Scheduled jobs survive restarts because of the durable delivery ledger shipped in v0.19: if a gateway crashes before a finished message is confirmed sent, the answer is recorded in state.db and redelivered on the next boot instead of vanishing (Hermes v0.19.0 release notes).
How do you keep a team of agents safe?
Three settings turn an agent team from "scary" to "trusting but verifying":
- Smart approvals on default. An LLM reviewer judges flagged commands instead of asking you to approve every unfamiliar shell command. You get bugged less, but you stay safe. You can also define deny rules for commands you never want Hermes to run — even under yolo mode — and
/deny <reason>teaches the agent what to do differently next time (Hermes v0.19.0 release notes). - Secrets in a password vault, not
.env. Bitwarden and 1Password sources load keys at runtime instead of leaving them in plaintext. Per-variable provenance tells you which vault each key came from (Hermes v0.19.0 release notes). - Per-profile secret isolation. Kanban workers and routed profiles only see their own profile's secrets, so a misconfigured researcher cannot read the writer's vault entries (Hermes Agent Docs — Multi-profile gateways).
For a deeper look at giving production access to AI agents safely, see our AI agent production access security guide.
What this means for you
If you are a solo operator, freelancer, or small business owner already drowning in repeatable work — research, drafting, follow-ups, simple landing pages — an agent team is the upgrade from "AI helps me when I remember to ask" to "AI runs parts of my business while I do other work." The Quicksilver release is the first version of Hermes where the multi-agent pattern does not force you to babysit: 0.9-second cold starts make it feel responsive, smart approvals cut the interruptions, live subagent transcripts cut the silence, and durable delivery means a crashed process does not throw away finished work. For a broader 8-level framework for scaling with AI agents, see our AI agent business scaling guide.
The reasonable starting point is not ten agents. It is three — a writer, a researcher, and a builder — configured in the steps above. Run that for two weeks, measure what you delegate, and only then add a fourth profile for a real new job you cannot cover with the first three.
FAQ
Q: What is an AI agent team? A: An AI agent team is two or more autonomous AI agents running on shared infrastructure but with isolated memory, skills, and instructions, so each agent specializes in one role (writing, research, coding) without mixing up its context with the others.
Q: Is Hermes Agent free to use for a multi-agent setup? A: Yes. Hermes Agent is open-source under the MIT license and self-hostable on a $5 VPS, your laptop, or a GPU cluster. You pay only for the model API calls each profile makes — the runtime and the profile/delegation/cron features are free (Hermes Agent Docs).
Q: How many AI agents can one Hermes install run? A: There is no hard-coded cap. Each profile is a separate home directory with its own config, memory, skills, and gateway state, and a single multiplexed gateway can route many channels to many profiles from one bot token. The practical limit is your machine's resources and how many discrete roles you actually need.
Q: What is the difference between profiles and subagents in Hermes?
A: Profiles are long-lived, isolated agents you configure once (writer, researcher, builder). Subagents are short-lived helper agents a profile spawns mid-task to parallelize a single job — you can watch them work live with tail -f on their transcript, and their output survives a process restart.
Q: Can I run an agent team for free with open models? A: Yes. Pair Hermes with a local open model via Ollama and each profile runs at zero per-token cost. The trade-off is hardware: you need a GPU with enough VRAM for the model you pick. Our free local AI agent setup guide walks through a fully zero-cost stack.
Q: How does the Quicksilver release make an agent team practical? A: v0.19 (Quicksilver) cut cold-start time-to-first-token by ~80% (from ~4.3s to ~0.9s), made profile-based message routing first-class so one bot can serve many agents, added live subagent transcripts, durable delivery so finished results survive crashes, and moved secrets into Bitwarden/1Password instead of plaintext files (Hermes v0.19.0 release notes).
Discussion
0 comments