When we talk about autonomous AI agents in business, we usually picture a single universal assistant. You load it up with sales instructions, tech support rules, post-writing guidelines, and database access. In practice, however, such a monolithic bot quickly chokes: the system prompt balloons to hundreds of lines, the model mixes up its tone of voice, and it starts replying to frustrated customers with ad slogans.
The major update of the open-source Hermes Agent v0.23.0 framework, codenamed "The Hearth Release," from the Nous Research team marked a paradigm shift. Having merged over 1,650 pull requests and commits, the project crossed 100,000 stars on GitHub and evolved from a solo executor into a coordinated digital headquarters where specialized subagents run shared team chats, work night shifts, and hand in their work by the time you pour your morning coffee.
From solo act to headquarters: the Hearth Release philosophy
The fundamental idea of the Hearth release is deep specialization. Instead of trying to teach one neural network to do everything under the sun, Hermes's architects propose dividing a company's operations among independent digital employees.
Each subagent receives a narrow role context, an individual set of system tools, and its own spending budget. Their work is coordinated in a single multi-user space (War Room), where bots see colleagues' messages, exchange results, and build on each other's output.
Subagent roles: Phoenix, Dragon, Kraken, and night watchman Owl
In the release configuration, the Hermes team offers a ready-made matrix of working roles:
- Phoenix (Creative Copywriter): responsible for generating ad creatives, catchy headlines, and marketing funnels. Phoenix isn't distracted by technical details, focusing purely on copy conversion.
- Dragon (Lead Nurturing Manager): instantly picks up new leads, generates personalized follow-up sequences, and keeps the conversation warm until the client reaches a purchase.
- Kraken (Claims Resolution Specialist): trained in conflict de-escalation. When a customer is unhappy with the service, Kraken calmly analyzes the complaint, checks company policy, and proposes a fair compensation solution without unnecessary emotion.
- Owl (Night Shift Watchman): a continuous night-cycle agent. While the whole team sleeps, Owl collects server logs at 5:00 AM, tallies overnight sales, compiles an analytical summary, and drops a neat digest into the work chat.
Unified Gateway architecture and cascading echo protection
To synchronize the actions of a distributed bot team, release 0.23 deploys a new lightweight Unified Gateway built in Rust and Python. It consumes less than 180 MB of RAM and streams messages across browser, desktop, and messengers over WebSockets with latency under 40 milliseconds.
When agents collaborate in the same chat, a specific risk arises: Infinite Chat Echo. If bot Phoenix writes "I've prepared the copy," and Dragon replies "Great, passing it to the newsletter," Phoenix might politely respond "Thanks, awaiting results"—and that dialogue will loop for hundreds of iterations, burning through your token balance.
In version 0.23, this problem is solved at the core architecture level:
- Dialogue depth limit: the
max_consecutive_replies: 3parameter strictly cuts off bot-to-bot chatter, requiring a trigger from a human or an external event. - Tool Budgets: each agent is forcibly assigned a financial cap per session (for instance, no more than $2.0 per task). Once the limit is reached, the agent gracefully pauses and alerts the operator.
- Editable Conversation: a human can delete an unnecessary bot message or tweak a prompt right inside the chat window without restarting the entire session.
Practical team configuration and launching the War Room
The team specification is described in a clean hearth_team.yaml file:
version: "0.23.0"
team_name: "OperationsWarRoom"
gateway:
host: "0.0.0.0"
port: 8080
cross_client_sync: true
agents:
- id: "phoenix_copy"
name: "Phoenix"
role: "Ad Copywriter"
model: "nous-hermes-3-70b"
tool_budget_usd: 3.0
system_prompt: "You are an ad copywriter. Write concise, punchy offers that hit pain points."
- id: "dragon_leads"
name: "Dragon"
role: "Lead Nurturing"
model: "deepseek-v4.1-flash"
tool_budget_usd: 1.5
system_prompt: "You are a lead manager. Reply to new leads within 60 seconds and nurture them to a call."
- id: "owl_watcher"
name: "Owl"
role: "Night Shift Auditor"
model: "nous-hermes-3-8b"
schedule: "0 5 * * *"
system_prompt: "You are a night auditor. Collect overnight sales metrics and compile a summary by 6:00."
rooms:
- id: "hq_chat"
participants: ["phoenix_copy", "dragon_leads", "owl_watcher"]
max_consecutive_replies: 3
Launching the team gateway is done with a terminal command:
hermes-gateway --config hearth_team.yaml --start-daemon
The Hearth Release clearly confirms that AI agent evolution has definitively moved out of the "watch the model answer jokes" phase into coordinated team automation for real business operations.

