Hermes Agent v0.19.0 is here, and this release name isn’t just marketing.
The Quicksilver Release goes straight at one of the most annoying parts of using any AI agent: the dead air before it starts responding. Hermes now cuts cold-start first-token time from roughly 4.3 seconds to 0.9 seconds, an improvement of around 80% across every platform.
That speed boost is the headline, but it isn’t the whole release. Not even close. v0.19.0 also brings smarter command approvals, direct secret loading from password managers, live subagent transcripts, more durable message delivery, profile-based routing, new model providers, deeper reasoning controls, and a much faster desktop experience.
The scale is honestly ridiculous: 1,065 pull requests, 2,245 commits, 3,300 issues closed, and more than 450 contributors since the previous major release window.
First Token in About 0.9 Seconds
The biggest everyday improvement is simple: Hermes starts talking a whole lot faster.
Cold-start latency reportedly dropped from about 4.3 seconds to around 0.9 seconds. Reasoning models now stream their thinking live by default too, so you’re seeing activity immediately instead of staring at a blank pause and wondering whether anything’s happening.
That matters more than a benchmark chart might suggest. Agents are used in loops. You prompt, inspect, redirect, approve, and keep moving. Shaving several seconds from every cold start adds up fast across a long session. It makes Hermes feel less like a remote process and more like a responsive desktop tool.
The Desktop App Got a Full Speed Pass
The desktop client received more than 20 performance-focused pull requests, and this wasn’t just cosmetic cleanup.
Streaming Markdown is now described as 14 times cheaper, large diffs are virtualized instead of rendered all at once, and switching between sessions should feel considerably snappier. These aren’t glamorous release-note bullets, but they hit the exact places where an agent interface starts dragging during real work: long answers, large code reviews, and sprawling session histories.
Hermes Desktop also gets a capabilities page, session colors, incremental Markdown rendering, and a pile of smaller interface improvements. None of those features will carry the announcement on their own, but together they should make the app feel noticeably tighter.
Smart Approvals Become the Default
Hermes can execute commands, so approval design isn’t some minor settings-page detail. It’s one of the most important parts of the product.
In v0.19.0, an LLM reviewer judges flagged commands by default. You can also define deny rules for commands you never want Hermes to run. That gives you a smarter checkpoint between the agent’s plan and the machine it controls without forcing you to blindly approve every unfamiliar shell command.
That’s paired with another major security improvement: Hermes can now load secrets directly from Bitwarden and 1Password. Credentials are read when they’re needed instead of sitting around in a plaintext .env file.
The release also includes a broader security-hardening pass and secret redaction for exported sessions.
Live Subagents You Can Actually Follow
Delegated subagents aren’t invisible workers disappearing behind a spinner anymore.
Hermes v0.19.0 lets you follow a delegated subagent’s transcript live. Its result also survives a restart, which makes longer-running delegated work far less fragile. You can watch what the subagent is doing while it works, then recover the completed result even if the main process gets interrupted.
That durability extends to final answers through the new delivery ledger. If a gateway finishes an answer but can’t deliver it to you, Hermes records the completed response and redelivers it on the next boot instead of quietly dropping it into the void.
For agents running through Telegram, Discord, Slack, WhatsApp, or hosted gateways, this is the kind of reliability work that separates a cool demo from something you can actually trust with recurring work.
Profile Routing Turns One Gateway Into Several Agents
A single Hermes gateway can now route different servers, channels, or threads to different profiles.
Each profile stays fully isolated, giving you a cleaner way to separate personal automation, development work, community channels, clients, or specialized agents without running a completely separate gateway for every one of them.
The release also adds durable cron audits, inline choice pickers on Telegram and Discord, WhatsApp polls, stacked slash skills, and a /model --once option for temporarily choosing a model without changing the session default.
New Providers, Models, and Reasoning Controls
Hermes adds Fireworks AI and DeepInfra as model providers, alongside support for newly listed models including GPT-5.6, Grok 4.5, Kimi K3, and Claude Fable 5.
Reasoning effort gets more granular too. New max and ultra effort tiers can be controlled per model, per Mixture-of-Agents slot, and per task. Instead of treating thinking depth as one global switch, you can spend more reasoning where it matters and keep routine jobs lighter.
There are also new /subscription and /topup commands for viewing your plan, previewing an upgrade, and applying it without leaving the terminal.
Sessions Can Leave Hermes Without Taking Your Secrets
Session history can now be exported as Markdown, HTML, or a Hugging Face trace format.
Exports support filtering and secret redaction, so it’s easier to archive work, publish reproducible agent traces, share debugging context, or move a session into another documentation pipeline without dumping credentials into the output.
It’s one of those additions that might look small on a feature list but becomes incredibly useful the second you’re documenting how an agent reached a result.
How to Update Hermes Agent
The project’s standard update command is:
hermes updateFor pip installations:
pip install -U hermes-agentThe Hermes Agent repository and its GitHub releases page have the source, tagged releases, and full changelogs.
Bottom Line
Hermes Agent v0.19.0 isn’t built around one giant new mode. It’s a systems release aimed at making the entire agent feel faster, safer, easier to follow, and harder to break.
The 0.9-second cold start is the feature everyone’s going to notice first. The smart approvals, password-manager integration, live subagents, delivery ledger, isolated profile routing, and durable exports are the features that’ll matter after the novelty wears off.
Quicksilver is the right name. Hermes didn’t just gain more abilities. The whole damn machine moves faster now.
