Time to first token drops by 80% in Hermes Agent

Nous Research's Hermes Agent reduces its turnaround time from 4.3 to 0.9 seconds. A major speedup for developer workflows.

Hermes version 0.19.0, dubbed The Quicksilver Release, revolves around a single obsession: latency. Nous Research estimates the time-to-first-token at startup to be around 0.9 seconds, down from 4.3 previously—a gain that applies to the command line, the gateway, the TUI, the desktop application, and scheduled tasks. Reasoning models stream their thinking live by default, and the response displays token-by-token rather than line-by-line.

The desktop application concentrates about twenty performance fixes: streaming no longer redraws the sidebar or tool rows with every token, large diffs no longer freeze the review panel, and switching between sessions no longer triggers layout recalculations. Custom endpoint settings have been added to the GUI, and session selector choices now survive restarts.

On the delegation side, each subagent now writes a readable log that can be followed live, tool call by tool call. Background-delegated tasks survive a process restart, and a delivery registry written to state.db ensures that a generated response no longer disappears if the gateway goes down before delivery confirmation on Telegram, Discord, or Slack.

Approvals are now controlled by default by a model that evaluates flagged commands, with each verdict applying only to the specific command examined, and refusal rules that apply even in yolo mode. API keys can move from the .env file to Bitwarden or 1Password. The Nous subscription can be managed from the terminal, a single gateway routes servers and channels to isolated profiles, 𝕏 search switches to grok-4.5, and Fireworks AI and DeepInfra join the supported providers.

The lab counts approximately 2,245 commits and 1,065 merged pull requests since the previous version, from over 450 contributors.