Hermes v0.20 "Herald": live voice, wake word, agent-to-agent calls

Key Takeaways

• Talk to Hermes — speech streams while the model thinks. Rather than waiting for a completed answer and then reading it aloud, the response is spoken as it generates, and speaking over it cuts it off and redirects. The video frames the resulting feel as closer to a phone call than to a prompt box.
• Wake word starts a session hands-free. Say the wake word from across the room and the agent begins listening; the demo triggers a chat session this way and gets a spoken reply without touching the machine. Practical use is leaving the desktop app or agent OS open on a second monitor.
• Agent-to-agent protocol, as a plugin. A new plugin implements A2A, so Hermes can discover other agents, be called by them, and orchestrate or be orchestrated in turn. The video singles this out as the item most worth watching: once agents can call each other, every skill and API wired into any one of them becomes reachable from the others.
• Grounded citation skill for research. A research mode that attaches sources to the claims it makes, positioned as a hallucination guard for deep-research-style answers rather than as a new model capability.
• Artifacts view with sandboxed live preview. The desktop app collects everything the agent has generated across platforms into one browsable view, and HTML output can be previewed inside a sandbox instead of being opened directly — the safer default for agent-generated code.
• Mid-turn correction. Type a correction while a turn is still running and the active turn redirects. This replaces the previous pattern of interrupting with /stop and re-explaining the whole task from the start.
• Smart approvals before risky actions. When the agent is about to do something it judges consequential — file deletion is the example given — it pauses and asks for confirmation first.
• Progress-aware compression. Context compaction now preserves recent conversation and tracks progress, so long sessions lose less of what just happened.
• Speed and scale of the release. First token at 0.9 seconds versus 4.3 seconds, with the telemetry gate cited as 54× faster, plus voice and text-to-speech improvements, core architecture work, security fixes, and bug fixes. The video credits 647 contributors on the release.

Commands & Code Mentioned

/stop

Mentioned as the command you no longer need in order to change course — mid-turn correction now redirects an in-flight turn without stopping and re-prompting. The update path shown in the video is the agent OS Manage → Update screen rather than a CLI command.