🤚 The Open-Palm Illumination

Nous Research has released Hermes Agent v0.17.0, and the update is so dense with features that the changelog reads like a SaaS company’s entire 2024 product roadmap compressed into a single Tuesday. The open-source AI agent — the one that learns from experience, builds its own skills, and models who you are across sessions — just gained background subagents, iMessage integration, a Blank Slate setup mode, and support for models with up to 1 million tokens of context. It also cut its cold start time by 19 seconds, which in AI agent terms is the difference between “productive morning” and “existential waiting.”

The headlining feature is delegate_task — a new call type that spawns non-blocking background subagents. These aren’t fire-and-forget scripts. They’re full agent instances that run in the background and feed their results back into your main conversation when they finish. Think of it as giving your AI assistant its own AI assistant, which, yes, is exactly the kind of recursive staffing decision that makes organizational charts cry.

On the communications front, Hermes can now send and receive iMessages via Photon Spectrum (no Mac relay server required), connect to WhatsApp Business Cloud API as an official Meta adapter, handle SimpleX groups with native attachments, and join the Raft agent-to-agent messaging network — a bundled platform where agents can communicate with other agents via metadata-only private channels. Your AI now has more messaging apps than your teenager.

👐 The Two-Handed Reality Check

The model roster alone deserves its own press conference. Hermes v0.17.0 adds support for z-ai/glm-5.2 with a staggering 1 million token context window, Anthropic’s Claude Fable 5 (yes, the one the U.S. government briefly suspended), Laguna M.1, Nemotron 3 Ultra, and — perhaps most intriguingly — Grok Composer 2.5 Fast, which is Cursor’s internal model accessible through an xAI OAuth subscription with a 200,000 token context window. Nous Research is essentially saying: we don’t care who made your favorite model, we’ll plug into it.

The desktop application received what can only be described as a dignity upgrade. macOS, Windows, and Linux versions now feature per-type notification toggles (so your agent can interrupt you for code errors but stay quiet about routine tasks), live subagent watch windows (so you can observe your background agents doing their background things), rebindable keyboard shortcuts, resizable terminal panes, and installable VS Code themes. It’s the kind of polish that makes you forget this project is open-source until you check the license.

Then there’s the Blank Slate mode — a new setup option that boots Hermes with exactly three things enabled: a model provider, file operations, and a terminal. Everything else — web access, browser, code execution, vision, memory, delegation, cron, skills, plugins, MCP servers — is explicitly disabled. The configuration writes to platform_toolsets.cli and agent.disabled_toolsets, which means your minimalist preferences survive updates. Nous Research built this for security-sensitive deployments, reproducible team setups, and teaching environments, but let’s be honest: some of us just want our AI agent to sit quietly and edit files without attempting to browse the internet, message our contacts, and schedule its own cron jobs.

🌿 The Gentle Awakening

What’s genuinely interesting about v0.17.0 isn’t any single feature — it’s the direction. Hermes is becoming an operating environment rather than a tool. Background subagents mean multi-agent orchestration without a framework. Raft means agent-to-agent networking without a centralized hub. The Skills Hub overhaul — which now includes previews and security scans before you install anything — suggests Nous Research is thinking about an app-store model for agent capabilities. And the atomic batch memory operations upgrade means the agent’s persistent knowledge layer can now handle multiple reads and writes in a single transaction, which matters when your agent is running five background tasks simultaneously and all of them want to remember something.

The curator optimization is worth noting too: routine runs — the periodic maintenance tasks Hermes performs to organize its memory and skills — now cost zero auxiliary model tokens. Previously, every routine run burned through your model budget even when nothing changed. It’s the kind of optimization that sounds boring until you realize it’s the difference between “I can afford to run this 24/7” and “I need to check my API bill before going to sleep.”

And then there’s the code verification system: after every edit, Hermes can now surface per-turn file-mutation summaries and language-server semantic diagnostics before proceeding to the next turn. Your agent will now pause to check if its code changes actually make sense — a courtesy that roughly 40% of human developers still haven’t adopted.

👑 The Crown Verdict

Hermes Agent v0.17.0 is what happens when an open-source project decides that competing with Claude Code, Codex, and every other AI coding tool isn’t ambitious enough — it wants to compete with your entire digital life infrastructure. It messages your contacts, delegates work to sub-agents, connects to agent networks, verifies its own code, manages its own memory, and now boots in 19 fewer seconds of your finite human time.

The Blank Slate mode is perhaps the most revealing addition. In a market where every AI product wants to do everything for everyone all at once, Nous Research shipped a feature that lets you turn almost everything off. It’s a concession to the reality that power comes from control, not from a longer feature list. The agent that grows with you, it turns out, is also willing to shrink.

Hermes Agent v0.17.0 is available now via pip install hermes-agent or through the desktop app on all major platforms. Your move.

Inspired by The new Hermes Agent update has me speechless…. by Alex Finn.

Your delegation strategy is showing. Orchestrate wisely.