Speaker A
Hermes's agent went quiet for two weeks. No big announcement, no flashy update video from the team, nothing on the usual channels. Just six release tags, one after another, from version 0.20.1 all the way to 0.20.6.
Detailed breakdown of six silent Hermes Agent updates revealing major features and improvements.
Ask about this video. Answers come from its transcript only — with the timestamp, so you can check them.
Generated from the transcript and can be wrong — check the timestamp.
Key Takeaways
- Hermes Agent underwent massive behind-the-scenes development with minimal public updates.
- The six patch releases prioritized stability and foundational improvements over flashy features.
- Several significant new features were quietly introduced, including multi-agent support and privacy-sensitive browsing.
- Users running Hermes Agent have already received these changes automatically via updates.
- A comprehensive, detailed changelog is forthcoming with the next major release.
What the video covers
- Hermes Agent released six consecutive patch roll-ups (v0.20.1 to v0.20.6) over two weeks with minimal public communication.
- Over 4,500 commits and nearly 2,000 pull requests were merged during this period, making it one of the most active development phases.
- The updates focused on stability for downstream users rather than curated feature announcements.
- Key new features include a multi-agent teammate system, web search without API keys, and consent-gated real profile browsing.
- The first release in this window was a large stabilization checkpoint with over 1,400 commits and 656 pull requests.
- Subsequent releases introduced foundational improvements like multi-gateway connection management and prompt caching for Light LLM Claude.
- Later updates added new integrations, improved security, and enhanced user experience across platforms.
- The team plans a full curated release (v0.21) with detailed notes and contributor credits covering this entire cycle.
- The video emphasizes the importance of these updates and recommends upgrading for security and new capabilities.
- The channel focuses on honest, hype-free analysis of open-source AI tools.
Chapters
- 00:00Introduction and Overview of Silent Updates
- 01:31Channel Introduction and Purpose
- 02:57Documentation and Community Impact
- 04:28First Release: Stabilization Checkpoint
- 05:58Second Release: Foundational Improvements
- 07:29Multi-Agent System Foundations
- 09:01Minor Fixes and Improvements
- 10:43New Skill Evaluator and Security Enhancements
- 12:04Web Search Without API Keys
- 13:41Additional Features and Update Receipts
Full Transcript — Download SRT & Markdown
Speaker A
And if you actually open any of those releases, every single one says the exact same thing. The full notes are coming later with version 0.21. So, for two weeks, thousands of changes went out into the world with almost no
Speaker A
explanation at all. Here is the part that should stop you in your tracks. Behind that silence, there were more than 4,500 commits, almost 2,000 merged pull requests, thousands of files rewritten, rebuilt, or replaced entirely. That is not a quiet couple of
Speaker A
weeks. That is one of the most active stretches this project has ever had, and almost nobody outside the repository actually saw it happen. So, I did the only reasonable thing. I went through all six releases myself, commit by
Speaker A
commit, feature by feature, so you do not have to dig through GitHub yourself to figure out what changed. And I promise you, some of what shipped in this window is a genuinely big deal.
Speaker A
There is a new multi-agent teammate system quietly taking shape. There is a way to get web search working with zero API keys right out of the box. There is a browsing feature that raises real questions about privacy and consent. And
Speaker A
there is a single release near the end of this window that is so packed with changes, it almost feels like it should have been its own separate version entirely. By the end of this video, you will know exactly what changed, why it
Speaker A
matters, and whether it is actually worth updating for. So, let us get into it. Before we dive in, if you are new here, this channel is all about breaking down open-source AI tools honestly. No hype, no sponsored fluff, just what
Speaker A
actually works and what is not. If that sounds like your kind of channel, go ahead and subscribe so you do not miss the next breakdown. And if this video ends up being useful to you, a like genuinely helps more than you would
Speaker A
think. And sharing it with someone who uses Hermes's agent might just save them from missing all of this the same way most people already have. All right, let us break down what actually happened. To understand why six releases in a row
Speaker A
shipped with almost no explanation, you need to understand what these particular tags actually are. They are not curated feature releases in the traditional sense. They are patch roll-ups built specifically for people running Hermes's agent through Docker images, hosted
Speaker A
deployments, or anyone who just installs from the latest tag and expects it to work. Their entire purpose is stability for downstream users, not storytelling for the community. Here is what that means in practice. If you have been running Hermes's agent through the update
Speaker A
command over these past few weeks, you have already absorbed every single change in this video without ever seeing a change log explain any of it. New features appeared on your machine. Bugs got fixed. Entire systems changed underneath you. And the only place any
Speaker A
of it was documented was buried inside commit messages and pull request titles that most people will never read. The team has said the full curated version is coming with the next major release.
Speaker A
And that version will apparently document everything going all the way back to the start of the cycle, complete with feature breakdowns and full contributor credit. That is a reasonable plan on paper. But for a project this size, with a community this active,
Speaker A
going dark on documentation for weeks at a time is worth talking about honestly, not just brushing past. So, that is exactly what we are going to do. I am walking through all six releases in the order they shipped, starting from the
Speaker A
very first one right after the previous major version, all the way through to the massive one that closed out this window. Some of these releases are quick and forgettable. Others are going to take a real chunk of this video because
Speaker A
of how much landed in them. Let us start at the beginning. The first tag in this window landed on August 13th, and honestly, this is the release with the least actually talk about, which makes sense once you understand what it was
Speaker A
for. This tag existed purely to roll up everything that had already piled up since the previous major version. We are talking about roughly 656 merged pull requests and over 1,400 individual commits. All bundled into one stable checkpoint for anyone installing fresh
Speaker A
or pulling the latest tag. The scale of it is still worth pointing out even without a detailed breakdown. Over 2,000 files were touched in this one release alone with more than 200,000 lines added and around 75,000 lines removed. Close
Speaker A
to 500 open issues got closed out in the process. That is an enormous amount of churn for what the team is calling a simple stabilization checkpoint. What actually got stabilized covers a wide surface area, too. The desktop application, the gateway platforms that
Speaker A
route requests between different providers, the installer scripts people use to get Hermes running in the first place, the internal tool system and the catalog of supported providers all received fixes and improvements in this pass. None of it was itemized
Speaker A
individually, which is frustrating if you like details, but it does tell you something important. This was the calm before a much busier stretch and it set the foundation for everything that came right after it. Let us move on to where
Speaker A
things actually start getting interesting. Three days later on August 16th, the next release landed. And this is where things started getting more interesting even though it still flies under most people's radar. Nearly 400 pull requests and just under 1,000
Speaker A
commits went into this one, and while none of it is flashy on the surface, a lot of it quietly makes everything else in this video possible. One of the bigger additions here is a multi-gateway connections registry on the desktop app.
Speaker A
If you have ever juggled multiple provider connections at once, switching between different accounts or different gateway setups, this is the kind of feature that prevents those connections from stepping on each other. Alongside it came profile scope refreshes, meaning
Speaker A
your different profiles can refresh their credentials and settings independently instead of everything getting tangled together. Health checks and deep links for MCP connections were also added, giving people a clearer, faster way to see whether their tool connections are actually working instead
Speaker A
of finding out the hard way mid-task. The command line side of things picked up several smaller, but genuinely useful improvements, too. Windows users got better update probing, so the tool can tell when a new version is actually available. Support for the Kitty
Speaker A
keyboard protocol was added, which matters if you use a modern terminal emulator and want your key presses handled correctly. And the chat command's continuation flag, the dash C option, got hardened against edge cases that used to cause problems. Perhaps the
Speaker A
most practical addition in this release is prompt caching for Light LLM Claude when it's routed through the OpenAI compatible wire format. If you are running Claude through tooling that expects an OpenAI style interface, this means repeated prompts can now be cached
Speaker A
properly, which translates directly into lower costs and faster response times for anyone using that setup. On top of that, cron job handling got more resilient, authentication resolution through profile scopes was cleaned up, and installer reliability improved on both Linux and Windows. None of this is
Speaker A
the kind of update that makes headlines, but this is exactly the sort of foundational work that keeps a fast-moving project like this one from falling apart under its own weight. And as we are about to see, the very next
Speaker A
release, which shipped the same day, is where things start getting a lot more ambitious. Here is where things get strange in an interesting way. This
Speaker A
[music] and yet it carries a completely different weight to it. Around 125 pull requests and 250 commits landed here, which sounds small compared to what came before, but the actual content inside this window punches way above its size.
Speaker A
The headline addition is something called bot mode, bundled in as its own plugin built around what the team calls a core team mate protocol. Strip away the branding for a second, and what this actually means is that Hermes agent
Speaker A
gained the foundation for multiple agents working together like teammates in a shared space, rather than everything being confined to a single assistant handling one conversation at a time. This is still early here, just the groundwork being laid, but it is a
Speaker A
meaningful shift in direction, and we are going to watch it grow across the next few releases in this video.
Speaker A
Alongside bot mode, this release also brought a migration to the newer MCP 2.x SDK, along with support for the stateless protocol version dated late July. If you are not deep into the protocol side of things, the short version is that this keeps Hermes agent
Speaker A
compatible with the latest standards the broader model context protocol ecosystem is moving toward, which matters a lot if you rely on third-party tool integrations. A new provider plugin called command code was added as well, giving users another option in the
Speaker A
provider catalog. For anyone using computer use features, meaning letting the agent actually control applications and interfaces, this release brought updated runtime contracts through quad driver 0.20, keeping that capability aligned with the latest driver version.
Speaker A
There's also a solid batch of reliability work under the hood. Python sub-process handling got isolated more carefully to avoid environment variable conflicts. The cron scheduler gained real self-healing abilities, recovering from file handle exhaustion, cleaning up stale job claims, and re-arming jobs
Speaker A
that had gotten stuck. Session handoff bugs that were causing data loss got fixed as well. Finally, a handful of smaller, but genuinely useful improvements were ported in from the broader ecosystem. Security scanning was added for plugin installs. New work tree
Speaker A
and rollback commands arrived with protections for hand-edited files. UTF-16 file reading was fixed, and tool call Idaho handling for Gemini 3 was corrected. Individually, these are small details. Together, they show a project actively absorbing improvements from everywhere it can find them. And with
Speaker A
bot mode now planted, the next release is where we start to see it actually take shape visually. Two days later, on August 18th, the next release landed, and this is the smallest window in terms of raw numbers out of all six we are
Speaker A
covering. Around 74 pull requests and roughly 146 commits went into it. But despite being the quietest release by volume, this one is actually the most visible if you are someone who cares about how the app looks and feels
Speaker A
day-to-day. The desktop application got a real visual overhaul here, built around a glass and translucency aesthetic. Think matte glass surfaces, a frost picker for adjusting the effect, and pre-selection support specifically tuned for macOS. It is a purely cosmetic
Speaker A
shift on paper, but it noticeably changes the feel of the app for anyone using it daily. More importantly, the sidebar got restructured into a tabbed layout separating sessions from bots, with the ability to hide or unhide individual bots as needed. This is a
Speaker A
direct evolution of the bot mode groundwork laid in the previous release, and it is the first time that system actually became something you can see and interact with directly in the interface. Alongside the sidebar work, several bugs in bot mode's group chat
Speaker A
functionality were fixed, including issues with long-running turns from individual members, problems with how markdown was rendering in group conversations, and routing issues when bots were operating across different machines. On the security side, this release introduced something worth paying attention to if you install a lot
Speaker A
of community skills. It is called the Nvidia skill evaluator, and it performs tier one advisory scanning whenever you install a new skill, checking for both licensing issues and potential security concerns before that skill runs on your machine. Given how much of the ecosystem
Speaker A
now revolves around installing third-party skills and plugins, this is a meaningful safety net to have built in rather than something you have to set up yourself. Rounding out this release, cron job media sending got more resilient with configurable timeouts,
Speaker A
support for manual run attachments, and better surfacing when a scheduled job misses its fire time. There were also fixes to how the session database handles its event loop thread and resource contention. The update command got more honest about branches that are
Speaker A
parked rather than actively tracked, and Kanban boards gained native operating system notifications. Small release by the numbers, but one that made bot mode real for the first time. Next up is where things start accelerating again, and fast. Tagged on August 19th, but
Speaker A
officially released 2 days later on the 21st, this next window is where the pace picks back up in a big way. Around 323 pull requests and roughly 746 commits landed here, making it one of the heaviest releases in this entire
Speaker A
stretch. And it is packed with the kind of features this channel genuinely gets excited about. The single biggest addition in this release is something called the keyless web tier. And if you have ever been frustrated by needing an
Speaker A
API key just to get basic web search working, this is going to matter to you.
Speaker A
It rotates across five different vendors with automatic failover built in, meaning a completely fresh install of Hermes agent now gets working web search out of the box with zero API keys required and zero setup on your end. For
Speaker A
anyone who has ever hit a wall trying to get an agent framework running without first collecting a stack of API credentials, this removes that barrier entirely. Bot mode continued maturing significantly in this release, too.
Speaker A
Group room threads were added, letting conversations branch and organize more naturally. Long conversations can now be folded into summaries automatically, cutting down on clutter. Avatars got a visual refresh with what the team is calling blob face designs. And you can
Speaker A
now drag and drop PDF and other file attachments directly into bot mode conversations, which brings it much closer to feeling like a complete multi-agent workspace rather than an early experiment. The command line interface picked up a genuinely satisfying batch of quality of life
Speaker A
improvements as well. There is now a fuzzy search picker for switching between models, meaning you can type a rough guess at a model name and still find it quickly. A command palette was added, triggered with control P, giving
Speaker A
you fast access to actions without memorizing every command. And the status command now surfaces much richer information than before. On the reliability side, this release introduced execution discipline and runtime stall guards, which came directly out of evaluation findings from
Speaker A
Compozea, an external group that stress tests agent frameworks like this one. In plain terms, this means the agent is now better at recognizing when it is stuck in a loop or stalling out and correcting course instead of quietly wasting your
Speaker A
time and API credits. A few more additions round this release out nicely. The update command now generates receipts, so you have a record of what actually changed. Fleet management gained a plan verification mode for previewing changes before they apply.
Speaker A
New work tree list and prune commands were added for managing multiple work trees more easily. An entirely free zero authentication provider called Open Code Free was introduced, giving you yet another way to use the tool without needing to sign up for anything.
Speaker A
Multi-question clarify was added, letting the agent ask several clarifying questions at once instead of dragging out a back and forth one question at a time. And on the performance side, bot mode now paints its interface before fully loading. Compositor spinners were
Speaker A
added for smoother loading states, and the React compiler was enabled across both rendering engines. Cron jobs even gained persistent memory now along with the ability to set reasoning effort on a per job basis. This is genuinely one of
Speaker A
the strongest releases in the whole window, and it lines up perfectly with everything this channel talks about when it comes to free accessible AI tooling.
Speaker A
But as big as this one is, the next release somehow manages to top it. This is it. Tagged on August 27th, this final release in our window is by far the largest of the six with around 525 pull
Speaker A
requests and an enormous 1,313 commits packed into a single stretch. If the previous release felt substantial, this one is on an entirely different level, and it closes out this two-week window with more changes than some entire major version ship with. Let's start with the
Speaker A
feature that raises the most questions, and rightly so. Hermes agent now supports consent-gated real profile browsing, meaning the agent can use your actual default Chromium browser profile when it browses the web on your behalf, complete with your existing logins and
Speaker A
sessions. On Windows specifically, this comes with a close with approval flow, requiring your explicit confirmation before the browser session ends. This is a genuinely powerful capability, letting the agent act on sites where you are already signed in without you having to
Speaker A
hand over credentials directly. But, it is also worth sitting with for a moment because giving an agent access to your real browsing profile, cookies, sessions and all is a meaningful trust decision.
Speaker A
And the fact that it is consent gated rather than automatic tells you the team is taking that seriously, too. Around that same feature, the desktop browser was pulled out into its own dedicated operating system window rather than being locked inside the main app. A
Speaker A
managed SSH remote update engine was introduced for pushing updates across multiple machines at once, alongside a fleet profile rail for managing several machine profiles from a single place.
Speaker A
Put together, these changes start to paint a picture of Hermes agent evolving into something closer to a fleet management tool for people running agents across multiple machines, not just a single local assistant anymore.
Speaker A
The provider ecosystem also expanded massively in this release. Over 50 vendor-hosted MCP servers were added to the catalog, all live verified, including names you will likely recognize like Cloudflare, Grafana Cloud, Better Stack, and Railway. This is one of the largest single jumps in
Speaker A
available integrations this project has seen, and it means the pool of services you can connect an agent to directly just got dramatically wider without you needing to configure any of it manually.
Speaker A
Performance and cost also got real attention here. Web search and web extract results now get cached with a time to live, meaning repeated queries do not need to hit external services again unnecessarily. A technique called lean tail compression became the default
Speaker A
behavior, which in plain terms reduces how much unnecessary context gets carried around during longer sessions, keeping things faster and cheaper. Tool search was upgraded to support multiple queries at once with proper word stemming, making it noticeably easier to
Speaker A
find the right tool even if you do not remember its exact name. On the security and reliability front, this release quietly fixed one of the more annoying long-standing issues for Mac users. App and operating system keychain encryption was added for stored secrets, which
Speaker A
means no more repeated keychain prompts every single time you launch the app. Updaters now pause running gateways gracefully through a control socket instead of forcefully killing entire process trees, which prevents abrupt interruptions mid-task. And installs managed through images or package
Speaker A
managers now refuse unsafe in place updates entirely, closing out phase three of a fix that had been tracked for a while. A last wave of smaller improvements rounds out this release nicely. Cron jobs gained durable incident acknowledgements and clearer
Speaker A
error messages when version mismatches occur. Slack integration picked up link on unfurling controls. Docker containers now share consistent identities across runs. Terminal environments became pluggable, letting different back-end implementations be swapped in. And the model picker grew once again, adding GLM
Speaker A
5.3 flash, a free tier of Minimax M3, and Minimax H3 Max for video generation, giving you even more free and low-cost options to experiment with. This release alone could have justified its own dedicated video, and it is the single
Speaker A
best argument for why calling this whole window quiet was never really accurate. Step back from the individual releases for a second, and a clear pattern starts to emerge across all six of these versions. Bot mode did not appear fully
Speaker A
formed. It started as a bundled plug-in with a core team mate protocol, barely visible outside of a change log entry.
Speaker A
Then it got its own place in the sidebar. Then it gained group rooms, folding summaries, drag and drop attachments, and real performance work to make it load faster. What began as a quiet experiment three releases ago is now a genuinely capable multi-agent
Speaker A
workspace sitting right inside the app. And it grew that way specifically because nobody was watching closely enough to notice it happening in real time. There is a second thread running through this entire window, too. One that speaks directly to the kind of
Speaker A
tools this channel tends to cover. The keyless web tier removed the need for API keys just to get basic web search working. The open code free provider gave people a completely free, zero authentication way to use the tool. And
Speaker A
by the final release, the model picker had grown to include free and low-cost options like Minimax M3 and GLM 5.3 flash. None of these arrived in a single dramatic announcement. They arrived quietly, one release at a time. And
Speaker A
together they represent a real shift toward making this tool accessible without requiring a stack of paid API subscriptions just to try it. And then there is the maturity thread, which might be the most important one of all.
Speaker A
Skill installs now get scanned for security and licensing issues before they run. Stored secrets can be encrypted through your operating system's keychain and instead of prompting you constantly. Updates now pause gateways gracefully instead of killing them mid-task. Browsing with
Speaker A
your real profile requires your explicit consent rather than happening automatically. None of these are flashy features you would put in a thumbnail.
Speaker A
But together they tell you this project is being built with real production concerns in mind, not just chasing the next feature to announce. So here is the honest question worth asking. Is six unlabeled patch releases in 3 weeks,
Speaker A
carrying nearly 2,000 merged pull requests between them, a sign of a project moving with serious momentum? Or is it a documentation problem that needs fixing before the next major version actually ships? I think it is genuinely both. The pace here is impressive by
Speaker A
almost any standard. But burying this much meaningful change behind a repeated placeholder message for weeks at a time makes it far too easy for even engaged users to miss things that actually matter, like a browsing feature that touches your real browser profile, or a
Speaker A
security scanner quietly protecting every skill you install. Speed without visibility eventually becomes its own kind of risk. And that is worth saying plainly rather than just celebrating the commit count. So, after everything we just covered, should you actually
Speaker A
update? If you are running Hermes Agent right now, and you are on anything before this window, the answer is yes, without much hesitation. The security improvements alone make this worth it.
Speaker A
From skill install scanning to keychain encrypted secrets to safer update handling that no longer kills your gateways mid-task. None of these are optional nice-to-haves. They are the kind of fixes you want running underneath you even if you never
Speaker A
interact with them directly. If you have an existing install, updating is as simple as running the update command from your terminal. If you're setting this up for the first time, the install script from the project's readme will pull the latest version automatically.
Speaker A
And thanks to the keyless web tier we talked about earlier, you will actually have working web search the moment it finishes installing, without hunting down a single API key first. A couple of things worth keeping in mind before you
Speaker A
dive in. Bot mode is still actively evolving. So, if you are trying it out, expect some rough edges, especially around long-running group conversations.
Speaker A
And if you are on Windows and plan to use the new browsing feature, take a moment to actually understand the consent flow before granting access to your real browser profile. It is a genuinely useful feature, but it is
Speaker A
worth using deliberately rather than clicking through without thinking about it. And that brings us to the end of this breakdown. Two weeks that looked quiet from the outside turned out to be one of the busiest stretches this project has had, and now you know
Speaker A
exactly what happened during it. If you found this useful, a like genuinely helps this video reach more people who are running this tool without realizing half of what we just covered. And subscribing means you will catch at the
Speaker A
moment those full curated release notes for version 0.21 actually drop, because you can bet we will be breaking that one down, too. Thanks for watching, and I will see you in the next one.
Topics:Hermes Agentopen source AIsoftware updatesmulti-agent systemweb search integrationprivacysoftware stabilityAI toolingsoftware developmentupdate analysis
![Ry and Julian – Their Story [The Last Sunrise] — Transcript](https://i.ytimg.com/vi/FVobJ6v8iNc/maxresdefault.jpg)
