Nous Research's Hermes Agent Pantheon Connects All Your AI Agents Into One Interface

Nous Research shipped Hermes Agent v0.21.0 with persistent multi-gateway connections for the desktop app. You can now connect the desktop to any combination of remote gateways simultaneously, whether those are Hermes Cloud instances or local agents running on other machines, and see every bot and chat across all of them from one interface.

The release carries the name Pantheon and consolidates the v0.20.1 through v0.20.6 infrastructure patches into one coherent minor version. The window covered roughly 5,800 commits, 2,475 merged PRs, and around 2,100 closed issues since v0.20.0.

How multi-gateway connections work

Previously, running Hermes across multiple environments meant juggling separate desktop profiles or SSH-hopping between machines to check on scheduled jobs. The v0.21.0 Connections registry treats each remote gateway as a persistent link the desktop maintains directly. The practical results:

• One desktop can subscribe to a Hermes Cloud instance, a home server, and a laptop agent simultaneously
• Every bot and chat from those gateways appears in a unified roster
• Gateways can go dormant when idle and coordinate drains for seamless deployments, useful for always-on services
• Profile-scoped refreshes, MCP health checks, and deep links keep the registry consistent as connections come and go

Bots Mode and agent-to-agent messaging

Every agent profile now gets a name, a deterministic avatar, and a slot in a shared roster. You can create group chats where multiple bots and human participants talk in one room, with @-mention support from the composer. The new hermes peer command lets any agent DM any other agent by handle, across profiles and gateways, from the CLI or from inside a conversation. Replies land in each agent's canonical Bot Chat, making exchanges between agents durable rather than fire-and-forget.

Subagents you can steer mid-flight

The delegate_task tool gained the ability to list running child agents, send course corrections to one while it runs, or stop it early and keep the partial result. Default limits were raised to 250 iterations and 10 concurrent children. JSON-schema validation can now be attached to child outputs, so a subagent returning malformed data gets caught before it affects the parent.

Scheduled agents received similar upgrades. Cron jobs now load and update persistent memory like any interactive agent, and a continuity=true flag carries each run's output into the next. A morning briefing job can deduplicate against what it already reported the day before instead of re-surfacing the same information.

Security changes

Writes to AGENTS.md, skills, and memory stores now require explicit write approval. A prompt-injected agent attempting to rewrite its own standing orders will hit the approval gate. A broader redaction sweep also closed secret-leak paths across terminal errors, .env reads, checkpoints, and ACP logs. The approval system now recognizes Windows destructive commands as well.

Installing or upgrading

Hermes Agent is open source with no separate licensing cost beyond whatever model providers you connect. Existing installs can run hermes update from the CLI. For a fresh install:

curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash

Several features that entered the branch during this window were reverted before shipping: Model Council mode (/council), the DCP context engine, and a WebSocket-only gateway server. FastAPI remains on the desktop boot path. Seq-stamped event replay for lossless desktop reconnect over WebSocket did make the cut, as did a roughly 50% reduction in default context usage.

Who benefits most

For anyone running a single Hermes profile on one machine, multi-gateway is more than needed. The feature earns its keep in the setup many power users have already built: a home server or VPS running always-on monitoring and cron agents, a work laptop running coding agents, and a Hermes Cloud instance for anything that needs to survive a reboot. Those were three separate contexts requiring independent checks. In v0.21.0, they appear as rooms in the same interface, and agents can communicate across the boundary directly.