Hermes Agent Quicksilver: 80% Faster Cold Starts, Durable Delivery, and Live Subagent Tracing
Hermes Agent v0.19.0 - codenamed Quicksilver - shipped on July 20, 2026. It is the largest release in the project's history by several metrics: 2,245 commits, 1,065 merged PRs, ~3,300 issues closed, and 450+ contributors over the window since v0.18.0. The headline change is an ~80% cut in cold-start time-to-first-token, from ~4.3 seconds down to ~0.9 seconds, applied across CLI, TUI, desktop, gateway, and cron.
Below that speed spine sits a stack of reliability and observability improvements: durable response delivery that survives gateway crashes, live subagent transcript streaming, smart approvals as the default, native password-manager integration, and profile-based message routing.
The latency fix
The "Initializing agent..." phase before v0.19.0 consumed ~4.3 seconds on every platform. Teknium's optimization (#59332, #59389) reduced this to ~0.9s - a 79% decrease. The fix applies everywhere the agent runs, with no configuration changes required.
"Hermes got dramatically faster - first token in a fraction of the time. If Hermes ever felt like it took a deep breath before answering, that breath is gone."
The desktop app received a parallel performance pass of 20+ targeted PRs. The markdown streaming renderer is now 14x faster (addressing a case where long replies consumed disproportionate CPU in the splitter), diffs are virtualized to prevent freezing the review pane, session switching no longer triggers layout thrash, and profile backends pre-warm on hover intent rather than blocking the cold-start path.
Reasoning models now stream their thinking live by default - no spinner while the model works through a 30-second reasoning chain. The response box paints per-token rather than per-line.
Durable delivery: responses that survive crashes
Before Quicksilver, if the gateway process died between generating a response and confirming platform delivery, that response was silently lost. The user never saw it, and the API call was already paid for.
The new delivery-obligation ledger (#67181) writes every final response to state.db before attempting platform delivery. On the next boot, any undelivered responses are re-sent. This covers Telegram, Discord, Slack, and every other messaging channel.
The same durability pattern extends to background subagent delegation (#63494). If a process restart interrupts a delegate_task job, results are restored through an ownership-checked ledger rather than discarded.
Live subagent transcripts
delegate_task dispatches now generate live transcript files that can be monitored with tail -f as soon as subagents launch (#67479). Each child agent gets one human-readable log containing every tool call, result, and streamed reply. The operator can fan out a fleet of subagents and watch any one of them work in real time without opening additional sessions.
Smart approvals as default
Hermes has always had a command-approval system, but the default was to ask the user for every flagged command. This created approval fatigue for heavy users.
Smart approvals (#62661) route flagged commands through an independent LLM reviewer. The reviewer categorizes each command:
- Safe commands are auto-approved
- Dangerous commands are auto-denied
- Uncertain cases escalate to the user
Each verdict is scoped to one command - a later command matching the same pattern gets its own independent review. The system pairs with user-defined deny rules (which block commands even under yolo mode) and a /deny [reason] command that explains rejections so the agent course-corrects on future attempts.
Password manager integration
API keys no longer need to live in a plaintext .env file. A new pluggable SecretSource interface (#59498) lets Hermes fetch secrets from Bitwarden and 1Password (op:// references) at load time.
Multiple vaults can be active simultaneously with deterministic precedence, conflict warnings, and per-variable provenance tracking. Future vault providers can be added as plugins. This consolidated eleven competing community PRs into one interface.
Profile-based message routing
A single multiplexed gateway sharing one bot token can now route specific guilds, channels, or threads to different Hermes profiles (#64835). Each profile has fully isolated config, skills, memory, and secrets. A work Discord server can route to a work profile while a hobby server routes to personal - all through one bot.
A second multiplex hardening wave (#65700) ensures one misconfigured profile cannot take down the entire gateway.
New models and providers
The model catalog gained GPT-5.6 (Sol/Terra/Luna + Pro variants), grok-4.5 (GA), moonshotai/kimi-k3, claude-fable-5, claude-sonnet-5, and tencent/hy3. Reasoning effort expanded with max and ultra tiers, selectable per-model with clamping on providers that support smaller scales.
Fireworks AI and DeepInfra joined as first-class providers, with Fireworks landing at the #2 slot in the provider picker. Upstage Solar was added via community salvage. LM Studio gained JIT model loading for local setups.
Terminal subscription management
/subscription and /topup commands let users manage their Nous plan without leaving the terminal. The flow shows current plan and remaining allowance, previews upgrade costs, and applies changes with scheduled-change banners and undo. The desktop app received a matching billing settings tab.
Session export
hermes sessions export now writes Markdown, Quarto, HTML, prompt-only, and Hugging Face-ready trace formats (#60186). The full filter surface includes age, workspace, and platform. An opt-in --redact pass scrubs secrets. Compacted-session lineage is stitched into one logical export.
The full scope
| Metric | Value |
|---|---|
| Commits | 2,245 |
| Merged PRs | 1,065 |
| Issues closed | ~3,300 |
| Contributors | 450+ |
| Files changed | ~2,465 |
| Cold-start TTFB reduction | -79% (4.3s to 0.9s) |
Quicksilver is available now via hermes update or by downloading the latest desktop build. The full release notes are on GitHub.
[^1]: Teknium. "Hermes Agent v0.19.0 (2026.7.20) - The Quicksilver Release." GitHub. July 20, 2026. [^2]: Simons, Tony. "The Quicksilver Release is here." X. July 20, 2026. [^3]: Buzovskyi, Ivan. "Hermes Agent Quicksilver thread." X. July 20, 2026. [^4]: Nous Research. "Hermes Agent Documentation." nousresearch.com.