a full week. Hermes Agent v0.21.1 landed as a rollup on monday, Hermes became part of Omarchy on tuesday, and I started a new series, Hermes Autopilot: prompts you paste into your agent and it does the rest. six new Wingtips, four short videos, four community builds and a Berlin meetup, and the 5090 was back on the bench every morning, so the local-ai section is back too.

this week's tips

six Hermes Wingtips this week, each checked against the live docs before posting:

• #65: hermes sessions import: carry a conversation from Claude Code or Codex CLI into Hermes Agent. hermes sessions import lists the sessions on your machine, newest first; pick one and it becomes a Hermes session you can resume. your messages, the answers and a short note wherever a tool ran come across; the other agent's files are only read. different from hermes import-agent, which moves your setup rather than the chat.
• #66: sessions.auto_prune: Hermes now tidies up ended sessions that have sat untouched for a long while, at startup, and gives the database space back when there is enough to reclaim. open, pinned and mid-conversation sessions stay put. hermes config set sessions.auto_prune false keeps everything forever; sessions.retention_days widens the window.
• #67: hermes debug share: one command builds the support report people always ask for: version and OS, model and provider, recent agent, gateway and desktop log lines, and which API keys are set (never their values). --local prints it in your terminal and nothing leaves the machine; the plain command uploads to a paste that expires on its own; --nous sends it to private storage only the Nous team can see.
• #68: /skill-name: every installed skill is also a slash command. /humanizer rewrite this reply in my voice loads the skill for that turn and runs the request with it; several /skill names stack in one message. for the agent to pick a skill on its own it reads a short index of every description, so a description that opens with when to use it gets picked up.
• #69: tool_loop_guardrails: inside one turn Hermes counts the same call failing twice, the same tool failing on new arguments, and a call returning the same result while nothing changed. at the count it writes a warning into the tool result so the model reads it and changes course. in the CLI, TUI and Desktop that is where it ends; gateway and cron runs also stop the turn, since nobody is watching. hermes config set tool_loop_guardrails.hard_stop_enabled true turns the stop on everywhere.
• #70: skills.external_dirs: a list of extra folders Hermes scans for skills, next to ~/.hermes/skills: a shared team repo, or a folder another agent tool already uses. every skill found there joins the index, hermes skills list and its slash command. on a name clash your local copy wins.

and one from my own Discord setup, one model per Discord thread via model aliases: hermes config set model.aliases.data <provider/model>, then @YourAgent /model data in the channel opens the thread on that model while every other thread keeps the default. I use my main model for thinking with me and a cheaper one for triage.

all 70 tips are now collected on one page on hermes-recipes, with a single llms.txt file you can point an agent at.

hermes logbook, between releases

• issue #8: Hermes Agent v0.21.1 shipped on monday as a rollup of 632 merged PRs. five you can use today: import Claude Code and Codex chats from the Desktop command palette; Settings, Appearance, Reopen Last Chat on Launch to open on a fresh chat instead; a per-job failure target for cron jobs (or local, disk only); subagents can hand a running background process back to the main chat, and results can arrive one by one as each finishes; and the write_file tool now tells the agent when a small edit would cost fewer tokens than re-sending the whole file.

shipped this week

two release explainers, a new docs round, and Hermes on a new OS:

• Hermes Agent is built into Omarchy: the Arch-based Linux setup where everything runs from the keyboard and one menu. Install, AI, Hermes Desktop puts the app on; Setup, Defaults, Agent, Hermes makes it the default agent for the terminal, the app and the menu entry. Omarchy ships its own omarchy skill into Hermes (keybindings, bar, hooks, themes), a second skill turns a crash notification into a diagnosis, and your Omarchy theme carries into the desktop app, the TUI and the CLI.
• web search with no key at all: a fresh install rotates web_search and web_extract across the public free tiers of Exa, Parallel, Firecrawl and Keenable, retrying on the next one if one throttles. any backend or key you add takes over immediately; web.keyless_fallback: false switches the ring off. for unthrottled service, a Nous Portal subscription runs both through the Tool Gateway with managed Firecrawl, or pin your own vendor key in hermes tools.
• live subagent view with steer and stop: workers now sit in the conversation itself, with what each is doing and for how long. steer sends a correction into a running worker without killing it; stop ends it early and the partial result still comes back. the same view in the CLI and in the Desktop app.
• docs update, round two: delegation defaults and output_schema written down, how state survives between execute_code calls, Bot Mode group chats that keep running after you close the Desktop, device-code MCP login for headless machines, one clear answer on automatic session pruning, and small fixes across the FAQ, Telegram, Windows and messaging pages.
• every release explainer on one page: seven so far on hermes-recipes, from real-profile browsing to Hermes on Omarchy.
• the feature-discovery answers, read back: 389 replies to my question on where you want to find Hermes features. the top asks: the agent surfaces the matching command or skill at the moment it becomes useful, a teaching mode that only mentions what you have not used yet, short per-feature clips, use-case templates, and a /whatsnew or /demo command as the simple version of all of it.

hermes autopilot, a new series

prompts you paste into your agent, which then does the work. the full prompt is always in the first reply under each post:

• #1: interview me: the agent asks you ten questions about how you work, saves what you said to its memory and shows you every entry. every later session starts already knowing it.
• #2: turn what you just did into a skill: paste it into a finished chat. it reads the session back, drafts a skill from the steps that worked, in order, with the pitfalls you hit, shows you the draft, and on your yes saves it as a slash command.
• #3: audit my own setup: in a fresh chat, it runs Hermes's two health commands, reads the report against how you actually use it, and hands you three changes with the exact line for each. read only; you run the commands.

on video

four short raw clips this week:

• Bot Mode from zero: make a bot in Hermes Desktop, give it a job, put it in a group chat with another bot, then remove it all again. you give it a name, a title and one sentence about what it does.
• the new Nous Portal, page by page: overview, models (Active Promos, Free Models, the searchable table with Portal price next to list price), the bundled tools, the Hermes Cloud page and the plans.
• Hermes on Omarchy in one minute: filmed on my old Dell. default agent from the menu, a question about Omarchy itself answered from the bundled skill, then a theme switch that the terminal and the app follow.
• comment mode in the Desktop browser: click the thing you want changed, type a note, stack a few pins, then Add comments. each pin lands in the composer as a cropped screenshot with the selector, markup and styles, so the agent knows exactly what you pointed at.

builders worth a look

• HermesWatcher's tokens-and-context guide: context, usage, cost and storage are four separate things and Hermes gives you a view on each: /context all, hermes prompt-size, /usage, hermes insights. a good place to start on those four.
• in-app tips in Hermes Desktop, by Brooklyn: the answer I liked most to the discovery question now exists. Hermes can point at one thing on screen with a small bubble while it talks to you, and the app surfaces the occasional tip of its own when you are idle.
• rlaope's oh-my-hermes: a layer on top of Hermes that decides what goes into each run, which model and effort, which skills, what counts as done. memory candidates go on a review card you approve, refuse or defer with the reason written down; Hermes's own memory is left untouched.
• thomasbek3's hermes-bot-kit: a bot's main chat shows messages only, tool calls and timers hidden; and when a bot starts working on a machine, a live view of that machine's desktop opens in a pane inside Hermes Desktop. plain files, no build step.
• professorpalmer's Puppetmaster: a swarm orchestrator for coding workers that picked Hermes Agent as its browser worker: headless, the browser toolset passed straight through, with a fallback to Hermes's local engine for pages on a private network.
• Tonbi's Desktop masterclass, part 2: 34 minutes on working in the app. my four day-one habits: one session per job, talk to it out loud when your hands are busy, hand long tasks to subagents, read the review pane before you accept.
• Hermes Power Users Berlin, 23 september: fifty Hermes users, a room in Kreuzberg, merch, an open floor. Nous supports community events; if you want to host one where you live, tell us.

local AI beyond Hermes

the RTX 5090 ran every morning this week. all on llama.cpp, greedy, same server:

• Qwopus3.8-27B-Flash against its base: the efficiency tune spends half the tokens and half the wall time on GPQA-diamond (4,229 vs 7,740 tokens per correct answer) and pays 8.6 points for it: 70.7 vs 79.3. it has learnt to stop early, and on the hardest questions it sometimes stops before it is done.
• a 125B MoE on a 32 GB card: Qwen3.8-Flash-Next, 79 GB file, experts in system RAM. MMLU 88.5, the highest on my board, five-task average 93.9, GPQA-diamond 54, 60 tok/s. from a 2-bit quant; the dense models it ties are 6-bit. the trade is prefill at 900 tok/s.
• the six llama-server flags, measured: on six models at a 16k prompt. --flash-attn changes nothing (auto already picks it), --parallel 1 is free speed, q8_0 KV cache is a memory lever on dense models only, q4_0 KV costs 17 to 29% of decode on every model, and the MTP draft head is worth 2.2x on Qwen3.8 and 0.7x on Nemotron Lightning.
• the reasoning budget curve: two thinking models given 4k, 8k and 16k tokens on the same 198 GPQA-diamond questions. Qwen3.8-27B goes 60.1, 70.2, 79.3. questions that finish inside the budget are 95 to 96% correct at every budget; all 101 lost points live in the capped ones. a 79% model with max_tokens at 8k is a 70% model on exactly the questions you wanted reasoning for.
• what a 16 GB card runs at 16k: 38 rows of the board turned into a fit table for 8, 12, 16, 24 and 32 GB. Qwen3.8-27B at 3 bits leads the 16 GB list (92.7 board, 12.8 GB peak) and beats every smaller model by 5 points. one thing to watch: the Q4_K_M file fits a 16 GB card, but its 16k context does not.

the data behind all of it lives in the rtx-5090-benchmarks dataset on HF and the llm-bench-rig repo.

that's the week. if you run Omarchy, Hermes is in the Install menu now. more Autopilot prompts and more bench mornings next week. see you next sunday.