a proper week this time. five new Wingtips, Bot Mode landed in Hermes Desktop, the community plugin scene had its best run yet, and the 5090 corner closed out a three-week benchmark arc and opened a new one. the local-ai section is at the bottom as usual.

this week's tips

five Hermes Wingtips this week, each verified against the live docs and run on my own boxes first:

• #47: cron prompts must be self-contained: a scheduled job wakes up with no chat history, so the prompt has to carry everything the task needs; a prompt written from inside a conversation usually doesn't.
• #48: platform toolsets: tool schemas can outweigh the system prompt itself; trimming one platform from the default 28 tools to the 10 it actually uses took the fixed per-call budget from 68.2KB to 18.1KB.
• #49: compaction thresholds: compaction fires at a share of your MAIN model's context window, 0.50 by default, with a raise-only floor for small-context models and per-model dials when you want different behaviour per model.
• #50: the [SILENT] marker: a cron job whose response contains [SILENT] delivers nothing; the output is still saved to the cron output folder, and a failed run always delivers.
• #51: checkpoints in one-shot runs: hermes chat --checkpoints -q snapshots the working directory before the first file change exactly like a live session, so even a headless CI run leaves a /rollback trail behind it.

shipped this week

• Hermes Logbook issue #4: the between-releases list, this time covering the run-up to v2026.8.13: /heartbeat, wake-word mic streaming, the anydoc document converter, the /learn knowledge-base shape and detached desktop updates.
• the Bot Mode docs page, section by section: a six-bullet menu of the new page, from creating your first bot to the teammate protocol.
• the CLI symbols glossary: a new docs page listing every symbol the Hermes CLI prints and what each one means.

around the ecosystem

the week belonged to Bot Mode:

• Bot Mode for Hermes Desktop: your profiles become teammates you can mention into any conversation, each with its own memory and tools; the official announcement carries a 90-second demo.
• everything Hermes Cloud unlocks: my rundown of what a hosted always-on Hermes gets you, down to the real daily price.
• skill installs now get an advisory scan: SkillEvaluator checks a skill before you install it, built on NVIDIA's benchmark of 300+ verified skills.
• live guided UI tours: a teammate demo of Hermes Desktop walking you through a website step by step in the preview pane, then touring its own composer; nothing scripted, it reads the screen live.
• Laguna S 2.1 stays free on Nous Portal: the free window got another two weeks.
• more ways to learn Hermes Agent: the learning map in one post: docs quick-links, the levelled learning path with time estimates, and the community masterclass embedded in the quickstart.

community corner

the part of the week I enjoy most, with a strong plugin-builder showing:

• 2,500 contributors: the hermes-agent repo crossed 2,500 GitHub contributors this week.
• Malt Radar: a whisky guide built with Hermes Agent: 4,000+ whiskies with spider charts, the data gathered by the agent, and the site's X account run by a Hermes cron six times a day in two languages.
• three deliberately incomplete agents: MASA gave a researcher, a builder and a verifier each a missing capability; the task routed itself across all three over Bot Mode and passed 10/10 verification with zero manual routing.
• Spotify from inside Hermes: Marco Franzon discovered the Spotify plugin and now picks his music without leaving his agent.
• Desktop plugins, heavily overlooked: @tonbistudio's video making the case that you just ask your agent for the plugin you want; no technical skills needed, only an idea.
• Hermes3D: the Claw3D author rebuilt around Hermes: a cosy 3D home where all your Hermes bots live and interact together.
• texting Hermes from an Apple Watch: after months of fighting third-party relays, one user had the built-in iMessage route working in about half an hour; his write-up is in the post.
• what would you remove from Hermes?: Teknium asked the community what to cut; answering is in your interest, because trimming buys everyone cheaper turns and faster first tokens.

local AI beyond Hermes

the 5090 finished the Qwen3.8-27B ladder arc and started a new one:

• the quant-tax ladder: seven GGUF rungs of Qwen3.8-27B on one RTX 5090; the whole spread from Q8_0 down to 2-bit is 2.9 points, and Q6_K ties Q8_0 at 93.7.
• what about longer contexts?: at 16k to 64k of context the picture splits: Q6_K retrieves 45/45 buried facts, the 3-bit file drops to 33/45.
• the hard benchmark tells a different story: the same eight cuts through GPQA-diamond; 4-bit and up is one flat band, and the 2-bit cuts pay a real cost the easy board couldn't see.
• the middle of the ladder, completed: the four remaining rungs through the buried-fact battery close out the depth column.
• the full-precision reference: BF16 ran quality-only over three nights of partial offload and landed mid-ladder at 93.5, next to Q6_K's 93.7; everything 4-bit and up matches full precision.
• thinking mode is the small quant's best friend: reasoning on is worth about 30 GPQA points, and the 11GB 3-bit file lands 2 points off the 27GB Q8_0, inside the noise band.
• Ornith 1.5 35B, day 3: 303 tok/s decode fully in VRAM, about 4 board points under the dense 27B reference and level with it on GPQA-diamond, still 258 tok/s at 32k of context.
• and with thinking on: Ornith goes from 52.0 to 81.8 on GPQA-diamond, level with the best dense rungs at three to six times their decode speed.

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 the checkpoints tip saves one of your CI runs, forward this to the friend still re-running theirs from scratch.