Hermes Agent

yeah their main branch should be a stable branch not a nightly (or minutely, lol) branch like it is now.

I get a lot of work done, that I could not do with other setups.

It’s also easy to just add my own pr to clean something up.

And the devs are easy to get into contact with imo.

They will need to brush up on their reliability very soon if they want to push into enterprise

I’ve had to debug quite a few things.

Use your Hermes agent to book you the accommodation. It's far more cheaper and less time consuming than what you would get from manually booking. I booked this hotel with my hermes agent for like $100 less than Booking_com the biggest hotel booking website apparently, lol.

It's far more cheaper and less time consuming than what you would get from manually booking.

I have it help me keep track of my studies, email, calendar and make sure linear is up to date. It's also allowed to fix stuff in my home assistant instance.

In my case anyways, Hermes starts session with shit ton of its instruction plus memory injection and list of all the tools available. It all adds up to longer prefill time before the llm model is able to start answering. It is just the nature of the agent unfortunately.

I ran 3.8 q3 and it is great - but it just doesn't work for hermes in my case. 80k context length is too short, and if I spill over to ram it becomes very slow. Which is what the problem is.

I think they do, I remember they did an announcement about that, but I don't like that they excluded CLI installations

I type "hermes update" in a terminal every so often. Usually before I do something substantial. I have several Hermes installs, and use them for various things, which I access via the desktop app, mostly. I have two that run on Linux servers I own, and then I have 3 machines I access those from. I have a local Hermes environment configured on two of the three, but rarely use them, it's mainly just so it's there if my connection to my VPN goes away, so anyway, i just run "hermes update" on any/all of them when I'm at the machine or connected to it so I'm generally current.

Idk, the desktop app’s “solution” for this is a button that shows how many commits were made since your last update and runs hermes update, unless they have a plan for actually fixing this

It gets around 40-45 tps base but it's actually hitting 100-110 due to dflash2 at 7 with 44-56% acceptance rate.

Runs extremely good, I use it everyday for different tasks (from managing my server stack to coding, personal assistant and so on).

stuck cards: hermes kanban notify-subscribe <task_id> --platform telegram --chat-id <id> sends you a message when a card completes or blocks. the Desktop Kanban plugin also raises a notification for blocked cards while the app is open, and there's a kanban_task_blocked plugin hook if you want something custom. same-family review: hermes kanban request-review <id> --reviewer <profile> moves a card into the review lane for a named profile, so your auditor on a different model family gets it as its own step.

part of this is built in. in config.yaml: `display:` `runtime_footer:` `enabled: true` `fields: ["model", "served_model", "latency"]` that adds the model and the turn time to the end of every final reply in Slack. served_model only appears when a fallback model answered instead of the one you configured. /footer toggles it from inside a chat.

tokens and cost aren't footer fields yet, but /usage in the same Slack thread shows input and output tokens with an estimated cost for that session.

I tried hermes 3:8b model but it still responds with the same errors

So - Hermes allows me to use local models (just upgraded to an M5 Max, 64GB) and mix in enterprise models from whoever, whenever i need more power… all without losing everything it knows about me or all the skills I’ve built for it over time.

Hooked it into DeepSeek flash and it was like a bulb went on and I saw what Hermes was actually capable of.

I tried a gemma 24B MoE and it made Hermes really useless with all sorts of issues.

I've noticed every update is breaking things more and more that Hermes has to repair since I'm using it on termux so I haven't updated since 2000 commits.

- The channel records live at hermes-assets.nousresearch.com/releases/channels/<name>.json — I probed them directly: stable.json, canary.json, and even main.json all return 404 right now. - Upstream issue NousResearch/hermes-agent#124309 (open, P2, filed Sep 29) documents exactly this: the stable release publication workflow is currently failing (at the Docker Hub promotion step), so the stable channel record never got published. Worse, the CLI lets you persist --set-channel stable even when the record doesn't exist yet, and there's no fallback for stable — only main has a built-in fallback to the git branch. - Net effect: main is the only source channel that resolves at all right now.

Self hosted and works great with Hermes.

I had downloaded Hermes before I tried this, and thought it didn't fit my needs (I want my conversations synched between my desktop, laptop, and phone, and I am not running a server) but I figured that since I had already installed it, I would see if I could tell it "make XYZ kind of wallpaper at 1920x1080" without having to deal with aspect ratios. While it was difficult to set up, I am happy with the results.

It sent much more detailed prompts to the image model than I would have sent and any resizing was taken care of automatically.

Long multi-step agentic workflows run day-to-day across all 3 models.

Good.

gpt-oss hits a hard 32K prompt wall (Hindsight prompts occasionally bounce over it).

YESSSSS, PEOPLE WHO ARENT USING KANBAN AND CRONs MAY AS WELL NOT BE USING HERMES AND GO BACK TO CODEX

There is nothing, and I mean nothing, that I have thrown at this that it can't do. I can't imagine spending hundreds of dollars on cloud tokens every month when this machine will do it all and do it fast from my house with no limits and with complete privacy

I can't imagine spending hundreds of dollars on cloud tokens every month when this machine will do it all and do it fast from my house with no limits and with complete privacy

Latency is my biggest frustration with Hermes

That worked thanks!

Also, Live Activities are kind of finnicky, first, clicking on them erases them, and second sometimes they disappear as the agent works for whatever reason, to make the live activity come back you have to go back in the app into the chat and let it update its status, kind of inconvenient if I want to check on long term work and see the live activity gone.

Hey I've been using the app for a day now, I had a problem where my tasks were getting cut off with "Operation interrupted" because iOS suspended Vory in the background, which drops the /api/ws socket, and Hermes fucking destroys detached sessions after a default 20-second grace (`HERMES_TUI_WS_ORPHAN_REAP_GRACE_S` in tui_gateway/server.py), so I fixed it by raising that env var to several hours in ~/.hermes/.env and restarting the dashboard service.

Fast both directions. MTP-4 ≈ +50–70% decode vs no-MTP; prefix cache hits ~94–96% on agent turns so repeated prompt text is reused, not re-prefilled.

Good. `enable-auto-tool-choice` + `qwen3_xml` parser handle multi-step tool calls cleanly.

I notice that the redde UI on my iPhone 16 is very sluggish/choppy for this longer thread (whereas it is smooth and snappy on a newly started thread).

Also - I'm noticing very rapid battery drain with this update. I can't know for sure why, but I suspect it is because it is now streaming all the thinking. If this theory is correct, you'll only notice rapid battery drain with models that think a lot such as qwen 3.8 27b on xhigh.

Mine wakes up my Tesla over WiFi at 7 pm every night and if I don’t have at least 50% in the battery it sends me a telegram message reminding me to go out and plug the car in

context compression failures from time to time

- No issues with the tools so far

Love you guys and your work. Keep building. I burn through roughly a bill tokens a day running multiple projects and HERMES is what has been helping me keep everything organized and sane all this time. Without it, i would have already become a mad man. Sky's the limit if you know how to use HERMES to its full potential.

Yes, doing hermes update on the laptop fixed things.

Basically I "journal" by doing voice notes on my phone, they get synced to my Mac, auto transcribed via a whisper job into a markdown file, and then I have a distillation job that adds [[wikilinks]], #topics, headers, removes filler words, etc. Then when I chat with my agent I effectively have a version of Karpathy's LLM wiki set up such that qmd searches my markdown notes for topics relevant to the conversation, details from notes and conversations are used to update entity files (the "wiki), and I get really good grounded perspective.

I've had a lot of success doing this and similarly use Hermes for self reflection. I build a skill around it called my "philosophical-sparring-partner" and my agent's soul is specificay configured to not blindly agree with me, to challenge assumptions, base perspective in facts and not inferences, etc.

piggybacking to also say Thank You. Hermes has been some impactful as someone with executive dysfunction. my head feels less cluttered because i know i can shoot off a text and let my assistant remember it. It's made a big impact so thank you.

Pretty snappy for day to day task. But if I require complex planning(for coding), I use glm5-3 on neuralwatt How well does it handle Hermes tools? Does everything. Doesn’t do a lot of reasoning and is very fast.

Works pretty much as an always-on assistant. Any offloading, swapping, crashes, or other limitations? Model has ngrams. No crashes.

It fast and smooth.

I also am having this issue. I can connect to my local backend, but not my remote backend.

How does it feel in everyday use? amazing How well does it handle Hermes tools? fine

How well does it handle Hermes tools? Well, barely have any issues. If I do it's because context has been compacted often

How does it feel in everyday use? Amazing. Not frontier-level, but it handles most things I throw at it and doesn't seem to hit any guardrails for reverse engineering. Since I usually let it do it's thing overnight it's even handling 5+ compactions well. How well does it handle Hermes tools? Well, barely have any issues.