Verdict: Hermes Agent v0.19 — the Quicksilver Release — is the most practical update the open-source agent has shipped to date. It does not add a flashy new model or a paradigm-shifting feature. Instead, it fixes the things that actually make agents frustrating to run day-to-day: slow startup, lost replies, approval fatigue, secrets sitting in plaintext, and subagents you cannot see. If you tried an AI agent before and gave up because it felt sluggish or fragile, this is the release worth revisiting.
Last verified: 2026-07-23 · ~2,245 commits · ~1,065 merged PRs · ~3,300 issues closed · 450+ contributors since v0.18.0
- Startup: ~80% faster first-token time (4.3s → 0.9s, vendor-reported)
- Security: Smart approvals by default; Bitwarden + 1Password secret sources
- Reliability: Durable delivery ledger so replies survive gateway crashes
- Visibility: Live subagent transcripts you can watch in real time
- Models: New providers (Fireworks AI, DeepInfra) + higher reasoning tiers (max, ultra)
Pricing/limits change often — last checked July 2026. Hermes Agent is free and open source under the MIT license; you only pay for model API usage.
Why the Hermes Agent 0.19 update matters
Hermes Agent, built by Nous Research, is an open-source autonomous AI agent that runs from your terminal, a desktop app, or as a gateway connected to Telegram, Discord, Slack, or WhatsApp. It uses tools, runs commands, remembers your work across sessions, and hands off jobs to subagents. The v0.19.0 release — published on GitHub on July 20, 2026 — is called the Quicksilver Release, after Hermes the messenger god. The theme is speed, but the real value is reliability. This update targets the friction points that make agents feel broken in daily use, rather than adding new capabilities you might never touch.
The scale is substantial: since the previous major release (v0.18.0), there were roughly 2,245 commits, about 1,065 merged pull requests, approximately 3,300 issues closed, and over 450 community contributors, according to the official release notes. That is the largest contributor window the project has seen. But what matters is what those commits actually fixed.
If you are already running Hermes, also see our guide on how to build an AI agent team with Hermes Agent for the multi-agent setup this update makes more practical.
How fast is Hermes Agent 0.19 now?
Hermes Agent v0.19 cuts cold-start first-token time by approximately 80%, from about 4.3 seconds down to roughly 0.9 seconds, according to Nous Research's own measurements. That reduction applies across every surface: the command-line interface, the gateway, the terminal UI, the desktop app, and scheduled cron jobs. Vous cannot fix a tool you are afraid to launch because it takes too long to warm up, and this update makes starting Hermes feel close to instant.
Beyond raw startup, the release added live reasoning streaming by default. Instead of staring at a spinner for 30 seconds wondering if the agent broke, you now see the model's thinking stream as it happens. The reply also paints token-by-token instead of line-by-line, so the interface feels alive instead of frozen. For anyone who has opened an agent, typed a message, and then watched nothing happen, this alone changes the daily experience. (Vendor-reported figures; no independent benchmark has been published.)
The desktop app received over 20 performance-focused pull requests in this window. Long replies previously consumed about 14 times more processing power in the text engine than they do now — the markdown rendering path was rebuilt to be incremental. Large code reviews used to freeze the review panel; that is fixed. Switching between chat sessions used to stutter; that is fixed too. If you tried the desktop app before and it felt janky on heavy conversations, v0.19 is the version to retry.
For context on how this fits into the broader agent landscape, see our self-improving AI agent guide which covers how Hermes's learning loop works over time.
How do smart approvals work in Hermes Agent 0.19?
Smart approvals are now the default in v0.19, meaning an LLM reviewer judges flagged commands instead of prompting you every time. When Hermes wants to run a command that normally requires approval, a second model evaluates that exact command and decides whether it is safe. Each verdict covers only that specific command — a later command that looks similar is reviewed independently, so nothing slips through on a pattern match.
This matters because approval fatigue is one of the most common reasons people abandon agent setups. If you have to manually approve every rm, every file write, every shell command, you end up babysitting the agent instead of getting work done. Smart approvals let you walk away while still maintaining a safety checkpoint between the agent's plan and the machine it controls.
You can also write your own deny rules — commands you never want Hermes to run, no matter what. These continue to block even in "yolo" mode (where all approvals are bypassed). There is also a new /deny command where you type the reason for denying something; the agent reads that reason and corrects itself going forward. This gives you both automated speed and hard limits in the same system.
For a deeper look at how to secure agents in production, our AI agent production access security guide covers the full permission model.
Can Hermes Agent pull secrets from Bitwarden and 1Password?
Yes. Hermes Agent v0.19 can load API keys directly from Bitwarden Secrets Manager and 1Password at process startup, instead of storing them in a plaintext .env file. This is one of the most important security improvements in the release for anyone running agents on shared infrastructure or multiple machines.
Here is how the Bitwarden integration works, per the official secrets documentation:
- You create a machine account in Bitwarden Secrets Manager and generate an access token.
- Hermes stores that single token in
~/.hermes/.envasBWS_ACCESS_TOKEN. - On every startup, Hermes calls the
bwsCLI to pull your provider keys (OpenAI, Anthropic, OpenRouter, etc.) into the environment. - Rotate a key in the Bitwarden web app and the change takes effect on the next startup — no editing config files on every machine.
The 1Password integration works similarly, using the op CLI with op:// references and either service-account or desktop-session authentication. Both secret sources can be enabled at the same time, with a deterministic precedence ladder: your .env file wins by default, unless a source sets override_existing: true (Bitwarden defaults to true so central rotation works). Hermes also labels each detected credential with its source, so you can see at a glance which vault each key came from.
The practical payoff: if you are running Hermes on a VPS, a Docker container, or across multiple machines, you no longer risk leaving API keys scattered in plaintext files. One vault rotation propagates everywhere. For teams, this is the difference between a hobby setup and something you would actually run in production.
What is the durable delivery ledger?
The durable delivery ledger is a new mechanism that ensures the agent's final response survives a gateway crash. Before v0.19, there was a real failure mode: the agent finished its reply, the gateway process died before the message reached the messaging platform, and the answer just disappeared. You had no way to know whether the agent failed to finish or finished but dropped the result in transit.
With v0.19, every final response gets written to a durable ledger (stored in state.db) around the send operation. If the gateway crashes mid-delivery, the response is redelivered automatically after the next gateway boot. This covers Telegram, Discord, Slack, and other supported channels, as documented in the release notes and detailed in the delivery ledger pull request.
This is not flashy, but it fixes a genuine trust problem. In a long-running messaging setup — where Hermes is sitting in Telegram and answering questions while you sleep — a lost reply is worse than a slow reply. It encourages duplicate requests, manual recovery, and less confidence in the automation. The ledger silently eliminates that failure mode.
Subagent results got the same durability treatment. When Hermes hands a task to a delegated subagent, the worker's results now survive a process restart. You can also watch the subagent's transcript live — every tool call, every result, every reply streams as it happens — instead of waiting for a final summary to appear.
How to update to Hermes Agent 0.19
Updating is straightforward. The command depends on your install method:
# If you installed via the one-line curl installer
hermes update
# If you installed via pip
pip install -U hermes-agent
The hermes update command auto-detects your install method (git installer, Docker, or NixOS) and runs the matching update path. It tells you what changed, backs up your config, and applies the new version. After updating, run hermes doctor to verify everything is healthy.
Post-upgrade checklist
After updating to v0.19, take these steps to get the most out of the release:
Enable live reasoning. Reasoning streams by default in v0.19, but if you previously disabled it, run
/reasoning showin your session. Watching the thinking stream makes it obvious when your prompt is going wrong — you catch a bad instruction before the agent has run five tool calls on the wrong assumption.Move your keys to Bitwarden or 1Password early. Migrating later is harder than setting it up now. Run
hermes secrets bitwarden setup(or the 1Password equivalent) and move your provider keys into a vault before you accumulate months of config in plaintext. Secrets Manager is available on Bitwarden's free tier with limits, per the Bitwarden docs.Write two or three deny rules. Pick the commands you never want Hermes to run —
rm -rf /, force pushes to main, dropping production databases — and add them as deny rules. Use the/denycommand with a reason so the agent learns from the rejection. These rules block commands even in yolo mode, so they are your hard floor.Start with one agent doing one job. Before you build a fleet of subagents, get one workflow working reliably. The speed and visibility improvements in v0.19 make it easier to iterate on a single agent's behavior, which sets a better foundation for delegation later.
Only crank reasoning to max on hard tasks. The release adds two new reasoning tiers —
maxandultra— available in addition to the existingnone, minimal, low, medium, high, xhighlevels, per the configuration docs. Settingmaxorultraon every task burns tokens on simple work. Set the global default tomediumand use per-model overrides or the/reasoningcommand for harder tasks. You can also set reasoning effort per auxiliary task (vision, web extraction, compression) independently.Export your session history. v0.19 lets you export sessions as Markdown, HTML, or Hugging Face trace format, with optional secret redaction. Your conversation history is useful material — for documentation, debugging, or training data. Export now and then; the redaction option means you can share traces without leaking credentials.
What new providers and models are in Hermes Agent 0.19?
The v0.19 release added two new inference providers:
- Fireworks AI — set
FIREWORKS_API_KEYin~/.hermes/.envand use--provider fireworks. Fireworks uses native slash-form catalog IDs (e.g.,accounts/fireworks/models/kimi-k2p6). The default endpoint ishttps://api.fireworks.ai/inference/v1. - DeepInfra — available as a custom OpenAI-compatible provider, or on DeepInfra's hosted Hermes-Agent instances which provision a dedicated VM with a pre-configured DeepInfra API key. DeepInfra offers access to open-weight models (DeepSeek, Qwen, Llama) with competitive pricing.
The model list also picked up new frontier models, including GPT-5.6, Grok 4.5, Kimi K3, and Claude Fable 5. You can switch models with hermes model or the /model command at runtime. For a comparison of how these frontier models stack up, see our Qwen 3.8 vs Claude Fable 5 vs GPT-5.6 comparison.
Profile-based gateway routing is another addition: one gateway can now route different channels (Telegram, Discord, Slack) to different profiles, each with its own config, skills, memory, and keys. This lets you run a single gateway process that powers multiple distinct agents — a personal assistant on Telegram, a team support agent on Slack, and a research agent on Discord — without spinning up separate gateway instances. Just be aware that isolation only holds if credentials, tool access, and routing rules are configured correctly per profile.
What this means for you
If you are a creator, solo operator, or small team using AI for daily work, the Hermes Agent 0.19 Quicksilver update removes the three things that make agents feel broken: slow startup, lost replies, and secrets scattered in plaintext files. Start with hermes update, move your keys into Bitwarden or 1Password, write a few deny rules, and let smart approvals handle the rest. Then export your session history — your own conversation logs are genuinely useful material for improving how you work with agents over time.
If you are building with open-source AI agents for the first time, this is a good release to start with. The live reasoning stream and subagent transcripts mean you can actually see what the agent is doing instead of trusting a spinner. Pair it with a free open-weight model on DeepInfra or Fireworks AI to keep costs near zero while you learn. For a full free setup, see our guide on building a free open-source AI agent with Hermes and Kimi K3.
FAQ
Q: Is Hermes Agent free? A: Yes. Hermes Agent is free and open source under the MIT license, built by Nous Research. You can run it on a $5 VPS, a GPU cluster, or serverless infrastructure. You only pay for model API usage if you connect a cloud model like OpenAI, Anthropic, or OpenRouter. Local models via Ollama or LM Studio cost nothing beyond the hardware.
Q: How do I update Hermes Agent to v0.19?
A: Run hermes update if you installed via the one-line curl installer, or pip install -U hermes-agent for pip installs. The update command auto-detects your install method and runs the matching path. After updating, run hermes doctor to verify everything is healthy.
Q: What is the smart approvals feature? A: Smart approvals use an LLM reviewer to judge flagged commands instead of prompting you every time. When Hermes wants to run a command that normally requires approval, a second model evaluates it and decides whether it is safe. Each verdict covers only that exact command, so a later similar command is reviewed independently. You can still write hard deny rules that block commands no matter what.
Q: Does Hermes Agent work with Bitwarden?
A: Yes. Hermes v0.19 can pull API keys from Bitwarden Secrets Manager at startup using the bws CLI. You store one bootstrap token in .env, and every other provider key comes from Bitwarden at runtime. Secrets Manager is available on Bitwarden's free tier with limits. The 1Password CLI (op) is also supported as a secret source.
Q: What is the durable delivery ledger? A: The delivery ledger records every final response to a durable SQLite store before it is sent to a messaging platform. If the gateway crashes mid-delivery, the response is automatically redelivered after the next startup. This eliminates the failure mode where the agent finishes an answer but the answer disappears because the gateway died before delivery completed.
Q: What reasoning effort levels are available in Hermes Agent 0.19?
A: Hermes supports eight reasoning effort levels: none, minimal, low, medium (default), high, xhigh, max, and ultra. You can set the global default in config.yaml, change it per session with /reasoning, set per-model overrides, and even set effort independently for auxiliary tasks like vision and web extraction. Higher effort gives better results on complex tasks at the cost of more tokens and latency.
Discussion
0 comments