What If Your AI Actually Managed Itself?
You know that frustration. You’re in the middle of a complex workflow, and your AI assistant just… stops. Waiting. Frozen while it processes some tedious subtask you wish it would handle quietly in the background.
Here’s the thing. The team at Nous Research just dropped Hermes Agent v0.17.0 “Reach” on June 19, and they’ve been busy. Since v0.16.0, the community poured in approximately 1,475 commits across 800 merged PRs, touching 1,693 files and closing over 300 issues along the way. That’s the work of 245 community contributors, not some minor patch notes.
Your Messages Finally Work Like They’re Supposed To
Remember wrestling with BlueBubbles bridges to get iMessage working? Those days are dead.
The new native iMessage platform plugin runs on Photon’s managed line pool using device-code OAuth authentication. No Mac relay. No QR codes to keep alive every few days. No bridge processes eating your system resources. It’s free to start, fully managed, and actually feels like something Apple should have built.
Speaking of dead bridges, the official WhatsApp Business Cloud API adapter now sits alongside the existing Baileys bridge. You get Meta’s first-party hosted path with no bridge process to maintain and no QR scan rituals. Meanwhile, Telegram messaging leveled up to Bot API 10.1 with rich messages, better formatting, and cleaner long-message handling, all enabled by default.
Multitasking That Actually Multi-Tasks
The headline feature is background subagents. Set delegate_task with background equals true, and Hermes dispatches a subagent that runs asynchronously, returning a handle immediately. The model keeps working while the subagent grinds away. When it finishes, the full result re-enters your conversation as a new turn.
You can watch this happen in real-time through live subagent watch-windows that stream delegated activity into its own pane in the matured desktop app. Speaking of which, the desktop experience now includes rebindable keyboard shortcuts, native OS notifications with per-type toggles, automatic RTL and bidi text direction, and a resizable VS Code-themed terminal pane. You can install any VS Code Marketplace theme directly. There’s a composer model selector with per-model presets and per-thread composer drafts so you never lose work mid-thought.
The Tools Got Smarter Too
The memory tool now handles atomic batch operations through an operations array. Add, replace, and remove edits apply together against the final character budget. What used to require multiple turns now collapses into one reliable operation.
For image generation, image_generate now supports image-to-image editing and transformation across all backends. Pass an existing image and prompt to edit it, using the same tool pattern as video_generate.
And if you’ve ever banged your head against cron syntax trying to schedule tasks, Automation Blueprints are your salvation. Pick a blueprint by name, Hermes asks for what it needs, and the same form renders everywhere: dashboard, CLI slash commands, TUI, messenger, and documentation catalog.
Skills Hub Reimagined
The Skills Hub got completely rebuilt. Connected hubs, a Featured section, full skill previews before installation, and security scans on each skill. Skills from trusted taps including OpenAI, Anthropic, HuggingFace, and NVIDIA are browsable now instead of dumped as a flat list.
New model support spans z-ai/glm-5.2 with verified 1M context through OpenRouter and Nous, anthropic/claude-fable-5, laguna-m.1, nemotron-3-ultra, and xAI Composer 2.5 via OAuth. The web dashboard includes a full profile builder to set model, skills, and MCP servers without editing config.yaml, plus a machine-wide global profile switcher to manage every profile from one view.
Privacy and Security That Doesn’t Sleep
The RAFT agent network gateway bundles an adapter that connects Hermes to Raft as an external agent through a wake-channel bridge. Here’s the critical part: wake payloads carry only metadata like event IDs and timestamps. Never message bodies. It’s privacy-by-contract design baked into the architecture.
Secure dashboard login now features hardened authentication where every token-required endpoint returns 401 behind the OAuth gate, websocket auth uses the served dashboard token, and a warning fires when a public_url override gets silently rejected.
Security hardening includes fail-closed behavior on missing gateway adapter policies, secret redaction in request debug dumps, blocking suspicious MCP stdio configurations before probe, operator environment scrubbing before cua-driver MCP launch, and bounding TodoStore content length and count.
Optimization That Saves Your Wallet
Curator cost optimization now runs a deterministic inactivity sweep that prunes stale skills by default. The LLM-powered consolidation pass is now opt-in only via curator.consolidate true or hermes curator run –consolidate. Routine background curation costs zero tokens. You get the cleanup without the bill.
The Takeaway
Your AI assistant should handle the boring stuff quietly while you focus on what matters. Hermes Agent v0.17.0 “Reach” was built by 245 people who understood that. Download it from GitHub today, and let your assistant finally pull its weight behind the scenes, not in front of them.
Do you want to get Hermes Agent on your own VPS? Find the best VPS providers for one click Hermes Agent installation here!