release week. Hermes Agent v0.21.0, the Pantheon release, landed on monday evening with 2,475 merged PRs from 760+ contributors, and Nous crossed its 100,000th contributor PR the next morning. most of this issue is me unpacking what shipped: Bot Mode in the Desktop app, the MCP command center, live subagent steering, multi-gateway Desktop, one-click local models. five new Wingtips too. no personal benchmark posts this week (the GPU is booked for a study starting tonight), so the local-ai section sits this one out.

this week's tips

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

• #59: the fallback chain: hermes fallback sets up backup provider:model pairs that Hermes switches to on a rate limit, a server error or an auth failure. history, tool calls and context carry over, and your next message goes back to your primary. pin a local model last and no provider outage can fully stop you.
• #60: /bg vs /btw: /bg hands a prompt to a separate background session and brings the result back when done. /btw answers a side question about the current conversation from a snapshot, so the running task never stops. both work on messaging platforms as well as in the terminal.
• #61: cron preflight: before a scheduled job runs, Hermes checks that the provider key resolves, the attached skills have what they need and the delivery targets are reachable. a job that fails is marked blocked_config, alerts you once and makes no model call, so it sits at zero cost until you fix it.
• #62: cron.model: scheduled jobs do not have to share a model with your chat. hermes config set cron.model <model> sets a fleet default, and hermes cron edit <job_id> --model <model> --provider <provider> pins one job. keep the flagship for conversations and let a smaller model do the routine rounds.
• #63: compression.context_timeout_seconds: when a long chat gets summarised in the background, this setting is the guard against the summary model hanging. after 120 seconds of silence Hermes tries one backup model, then skips the summary, keeps your messages and tells you. it only counts silence, so a slow model that is still writing gets all the time it needs.

two longer ones for people setting up:

• Bot Mode fleet setup in five steps: my own setup is a small team, one thinks with me on Discord, one runs the schedules on a Mac Mini, one benches models on a GPU box. since v0.21.0 that pattern is built into the Desktop app: New Agent in the Bots tab, a model per bot, a group chat, routines attached to the bot responsible for them.
• new to Hermes Agent? start here: four pointers, the FAQ with 30 community questions answered, the docs, Tonbi's Masterclass video series, and the Wingtips search index.

hermes logbook, between releases

• issue #6: the PRs that landed between v2026.8.21 and v2026.8.27. web_search and web_extract results are cached for 20 minutes so a subagent fan-out stops re-billing you; 44 more services joined the MCP catalog via hermes mcp install; delegate_task workers now start with your AGENTS.md in their system prompt; the in-app browser grew tabs; and the Desktop can update your remote SSH boxes from a Managed updates section.

shipped this week

the Pantheon release, and what I made of it:

• Hermes Agent v0.21.0 is out: 2,475 merged PRs from 760+ contributors since v0.20.0. Bot Mode is built into the Desktop app, every profile becomes a named agent, you put them in group chats and @-mention any of them. around it: hermes peer for bot-to-bot DMs, cron jobs with continuity, the MCP command center, a CLI wave, and the agent now drives the in-app browser itself.
• the small terminal upgrades: Ctrl+P opens a command palette with every slash command searchable, the status bar shows live cache-hit rate and tokens per second, and hermes pets puts a pixel companion in your terminal. it does nothing for productivity, and that is the point.
• the MCP command center: one Desktop page where your servers and the whole catalog live together. adding a server is a paste, background health checks nudge you to re-auth when needed, a cost overlay shows what each server's tool schemas cost you in tokens with 30 days of usage, and hermes:// deep links install a server in one click behind an explicit confirmation.
• live subagent steering: the parent agent can now list its running delegate_task children, steer one with a course correction that lands at its next step without killing the in-flight work, or stop one and still get the partial result. you can also hand a child a JSON schema its answer has to match; defaults are 250 tool-calling turns per child and up to 10 children at once.
• bot-to-bot messaging, how it works: every bot carries the live roster of its teammates, the message_agent tool validates the target and delivers into that bot's own Bot Chat with a "Message from" attribution, and typing "@ researcher have a look at this" makes the active bot compose its own handoff. hermes -p <bot> chat does the same headless from the CLI.
• persistent multi-gateway connections for Desktop: my Desktop runs against a Hermes Cloud instance and can now hold every other Hermes box I own at the same time. Settings, Gateways, Add connection, name it, click Test and wait for "Reachable". switch machines from the Sessions sidebar; agents keep working in the background while you look at another one.
• Hermes Desktop sets up local models in one click: it reads your hardware, picks the model that fits, downloads it and configures the runtime. picking the right local model has been its own skill until now; this removes the biggest barrier to running one.

on the Nous Portal side:

• Fable 5.1 is live in Nous Portal at 20% off the API price. /model anthropic/claude-fable-5.1 --once runs it for a single turn, --global makes it your daily model. I asked for feedback on how it behaves inside Hermes: tool calling, long runs, instruction following.
• Gemini 3.8 Flash is on Nous Portal: a flash-class model is the workhorse of an agent setup, a fast daily driver, the auxiliary model for compression and web reading, or the engine for a fleet of subagent workers.
• the Nous Portal referral programme: anyone joining through a referral code gets $15 off a Plus plan and the sharer gets $10 in credits. my code is in the post if you want the discount.
• half price on any Nous Portal subscription through 9 september, code in the quoted announcement. still live when this lands in your inbox.
• the Hermes Agent FAQ passed 800 bookmarks and we are collecting questions for the next edition. if there is something you still do not have a clear answer to, reply to that post or to this email.

builders worth a look

• hooeem's "How to become a Hermes Operator" course: eight phases, 53 sections, on X. if you follow along, the Nous stack covers every step: Portal for models and the hosted tool gateway, Hermes Cloud for the always-on agent, Desktop as the control panel.
• Tonbi's Hermes Desktop masterclass, part 1: the author of the Hermes Agent Masterclass is doing the same for the Desktop app, three parts, the first covering install, the UI, settings and the different backends.
• vaelkeep's hermes-paper-agent: a cron that writes an actual newspaper overnight, front page, sections, page numbers. the numbers never come from the model (weather, markets and the household ledger are plain Python desks), and a checker lints every page before anything gets printed. repo with a live sample edition in the post.
• MrR0b0t's hermes-zvec-memory: a memory provider where your agent's long-term memory is a folder of Markdown files on your own machine. every recall cites the exact file and line, search runs keyword and meaning matching at once, and it plugs into the standard hermes memory setup slot alongside the built-in memory.

that's the week. the GPU box goes back to work tonight, so expect benchmark numbers in the next issue. if you are on v0.20.x, hermes update is worth your five minutes. see you next sunday.