# Nous Research ships Hermes Agent Bot Mode for multi-agent group chats > Source: > Published: 2026-08-31 21:16:10+00:00 # Nous Research ships Hermes Agent Bot Mode for multi-agent group chats **Version 0.21.0 adds persistent bot-to-bot messaging, steerable subagents, browser control and scheduled jobs that remember prior runs.** By [Ryan Merket](/author/ryan-merket) ยท Published Primary source: [Nous Research on X](https://x.com/NousResearch/status/2094515104670715940) ## Why it matters Hermes Agent is turning multi-agent systems into an inspectable product: persistent bot chats, remembered routines and live intervention replace hidden chains of agent calls. [Nous Research](https://x.com/NousResearch/status/2094515104670715940) released Hermes Agent v0.21.0 on August 31st, turning its desktop app into a shared workspace where users can create named AI agents, place several of them in group chats and let the bots message each other directly. [https://x.com/NousResearch/status/2094515104670715940](https://x.com/NousResearch/status/2094515104670715940) The release was directed by [Ryan Teknium (@Teknium)](https://x.com/Teknium), the Nous researcher credited with leading Bot Mode, the new Model Context Protocol management page and the agent's expanded memory for scheduled tasks. Teknium was also the lead author of Nous Research's Hermes 4 technical report, published in 2025, and has remained the central maintainer behind the agent runtime. The v0.21.0 notes credit him with roughly 1,340 merged pull requests during the release window. Nous Research calls v0.21.0 the "Pantheon Release." The name is grander than the interface change: Hermes Agent now presents a collection of autonomous processes as something resembling a team chat. Each agent profile receives a name, an avatar and a place in a shared roster. Users can open rooms containing multiple bots, mention a specific agent and inspect exchanges between agents instead of leaving delegation buried in logs or terminal output. The accompanying `hermes peer` command handles direct bot-to-bot messages across profiles and gateways. Replies are stored in each agent's canonical chat, making handoffs persistent and reviewable. A research agent can pass findings to a coding agent, for example, while preserving the exchange for the operator. ### Scheduled agents gain continuity The more consequential work sits beneath the chat interface. Scheduled Hermes jobs can load and update persistent memory, while a new `continuity=true` setting carries output from one run into the next. That allows a monitoring agent to compare new results against earlier reports instead of rediscovering and repeating the same information on every scheduled run. Each job also receives a durable notepad. A monitor mode can skip a model call when the underlying material has not changed, reducing unnecessary inference use. Outputs can be delivered into a bot's regular chat, where the bot can continue the conversation rather than treating the scheduled run as an isolated alert. Nous Research also made delegated subagents controllable after launch. Operators can list active children, redirect one while it is working or stop it and keep the partial result. Hermes Agent can validate a subagent's response against a JSON schema and report the cost of an individual delegation. The default ceiling rises to 250 iterations and 10 concurrent child agents, according to the release notes. Those changes move [Hermes Agent](https://github.com/NousResearch/hermes-agent) away from the familiar one-user, one-assistant session. Nous Research is building an operator console for persistent collections of agents, with chat serving as the control surface. The [avatars](/models/fal/veed-avatars-1) are presentation. Durable messages, shared history and mid-task intervention are the coordination layer that makes the presentation useful. ### The agent can operate its own browser Version 0.21.0 lets Hermes Agent navigate, click and read pages inside the desktop app's browser. The prior browser could display material without giving the agent direct control over the embedded tab. Users can now watch Hermes move through documentation or inspect a web application while keeping the work inside the desktop client. Nous Research also consolidated Hermes Agent's Model Context Protocol servers and catalog into a desktop command center. The page supports imported server configurations, background health checks, reauthentication warnings, estimated schema-token costs and 30-day usage by server. `hermes://` links can initiate an MCP installation after an explicit confirmation. The runtime remains model-agnostic. Its documentation lists Nous Portal, OpenRouter, OpenAI and self-hosted endpoints among the available routes, with model switching handled from the `hermes model` command. The new release adds six provider integrations and allows users to override model metadata such as context windows, pricing and capabilities without waiting for another Hermes update. ### Security controls meet a larger attack surface Nous Research says the release now requires approval before Hermes Agent writes to protected instruction files, including `AGENTS.md` , skills and memory stores. The update also expands secret redaction across terminal errors, environment-file reads, checkpoints and logs. Those controls matter as Hermes gains browser access, persistent memory, third-party tools and conversations among agents. Hermes Agent's own [security policy](https://github.com/NousResearch/hermes-agent/security) draws a harder boundary around those protections. It describes approval checks and redaction as heuristics rather than containment, and says operating-system isolation is the only security boundary against an adversarial model. The policy recommends wrapping the entire agent process in a sandbox when it consumes untrusted web pages, email, shared channels or MCP responses. Bot Mode makes the product easier to operate; it does not reduce the privileges granted to the underlying processes. Nous Research says the release covers about 5,800 commits, 2,475 merged pull requests and 2,100 closed issues since v0.20.0, with work from more than 760 contributors. Those are project-reported figures, and the version also rolls up infrastructure patches released as v0.20.1 through v0.20.6. Existing users can install it with `hermes update` .