Files
waggle-os/docs/addictiveness-audit-2026-05-28/BENCHMARK-hermes.md
Oleg Maslov 0c3e2ead3b
Some checks failed
Installer Smoke / installer-smoke (push) Has been cancelled
moving
2026-09-02 10:10:29 +02:00

78 lines
6.9 KiB
Markdown
Raw Permalink Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# BENCHMARK — Hermes Agent (Nous Research)
_Audit date: 2026-05-28. Verified via WebFetch on hermes-agent.nousresearch.com + github.com/NousResearch/hermes-agent + 6 third-party reviews._
## 1. Reality check
**Real and live.** Released **2026-02-25**, MIT-licensed, by **Nous Research**.
GitHub: <https://github.com/NousResearch/hermes-agent>. **Current version v0.14.0 ("Foundation Release"), 2026-05-16. ~170k stars** (95.6k at the 7-week mark — fastest-growing agent framework of 2026).
Companion repo `hermes-agent-self-evolution` adds DSPy + GEPA optimization. Docs site: hermes-agent.nousresearch.com.
**Corrections to your second-hand summary:**
- Release was **Feb 25, 2026**, not just "Feb 2026" — confirmed.
- Gateways are real, but list is **Telegram, Discord, Slack, WhatsApp, Signal, Email, CLI** (Email included; you missed it).
- "Sub-agents that spawn in parallel" is correct — isolated subagents with five sandbox backends (Docker, SSH, Singularity, Modal, Daytona).
- Platforms: **Linux/macOS/WSL2/Termux native; Windows PowerShell is early beta**, not first-class. Single-curl install only on Linux/macOS/WSL2.
- It is **not pure-OSS — Nous Portal is a paid hosted add-on** (300+ models + Tool Gateway: Firecrawl search, FAL image-gen, OpenAI TTS, Browser Use). Self-hosting works without it.
## 2. Daily-driver features (what locks users in)
- **Built-in learning loop.** Every ~15 tool calls Hermes pauses, analyzes what worked, and writes a reusable skill to `~/.hermes/skills/`. Skills self-improve on subsequent runs. This is the headline differentiator and visible in user reviews as the "compounding value" hook.
- **Persistent memory + Honcho dialectic user model.** FTS5 search over all past conversations + LLM summarization; an explicit model of "who you are" that survives across sessions and platforms.
- **`agentskills.io` open skill standard** — portable skill marketplace already nascent (`awesome-hermes-agent` repo lists community skills).
- **Natural-language cron** — "Send me a daily project digest at 8am to Telegram" parses to schedule + delivery channel.
- **Multi-gateway presence.** One agent reachable from 6 chat surfaces + email + CLI. Same memory across all.
- **Server-resident, not laptop-resident.** "Talk to it from Telegram while it works on a cloud VM" — runs on a $5 Hetzner VPS.
## 3. Addictive / sticky design
- **Variable reward via visible skill growth.** Users see `~/.hermes/skills/` directory fill up — concrete artifacts of "the agent got smarter today." Several reviews flag day-30 as the inflection point.
- **Daily trigger via cron + chat push.** Hermes initiates conversations (digests, briefings) rather than waiting to be opened — the gateway delivers to apps users already check.
- **Investment hook (IKEA effect).** Every interaction trains the user-model and creates skills the user "owns." Switching cost compounds invisibly.
- **Channel ubiquity.** Telegram/WhatsApp/Signal means engagement happens in the same threads where users already live — no separate app to remember to open.
- **No premium gate on the loop.** Memory + skills + cron are 100% free OSS, so the addictive layer is not behind a paywall.
## 4. Onboarding / hook moment
- **Install:** one curl line → `source ~/.bashrc``hermes setup``hermes`. ~2 minutes on Linux/macOS.
- **Day-2 hook (per reviews):** run **one** narrow repeating task (daily report, log triage) — once a skill auto-generates and triggers on day 2 via cron, users see the loop pay off concretely.
- **Day-30 inflection** is the most-cited retention milestone — the skills directory and user-model carry visible weight by then.
- **Friction:** CLI/SSH-first today. Issue #10488 tracks a "secure first-run web onboarding wizard" — they know non-technical users bounce. Not yet shipped.
## 5. UX surface
**Server-first, chat-on-top.** Daemon runs on a VPS; the user interacts via:
- **TUI** — full terminal interface with multiline edit, slash-command autocomplete, history.
- **Messaging gateways** — Telegram/Discord/Slack/WhatsApp/Signal/Email (cross-platform conversation continuity).
- **No desktop app, no native GUI, no browser app.** This is the biggest UX gap vs. Waggle.
## 6. Pricing / OSS vs hosted
- **Core agent: free, MIT, self-hosted.** No seat/feature paywall.
- **Infra:** $4$25/mo VPS + $2$15/mo LLM API → realistic floor **~$6/mo** (Hetzner + DeepSeek V4 with caching).
- **Nous Portal (optional):** subscription gives 300+ models routed + Tool Gateway (search/image/TTS/browser). No public flat price; per portal.nousresearch.com it's a managed sub.
- **Third-party managed hosting:** $6/mo (OpenClaw Launch) → $59/mo (FlyHermes); enterprise "PTG" tiers $5K$40K+ one-time.
## 7. Top 3 weaknesses Waggle can exploit
1. **No real GUI.** TUI + chat-bots only. No file browser, no canvas, no spatial workspace, no visual memory view. Waggle's desktop OS metaphor + Room canvas is a category-different surface for the ~80% of knowledge workers who don't live in terminals.
2. **Windows is second-class.** Native PowerShell support is "early beta"; the curl-installer flow is Linux/macOS/WSL2. Waggle ships Windows binaries as first-class.
3. **Non-technical onboarding is unsolved** (their own issue #10488). VPS + DNS + SSH is a hard wall. Waggle's local-first Tauri install is one-click — no infra to provision.
## 8. Top 3 strengths Waggle should match
1. **Visible compounding value.** The `~/.hermes/skills/` directory is the killer artifact — users *see* the agent get smarter. Waggle has Wiki Compiler and an Evolution subsystem, but neither surfaces as a "look how much I taught it" trophy case. **Action:** ship a Skills/Memory growth tile on the OS dock with delta counters ("+3 skills, +47 frames this week"). Mission Control tile is the natural home.
2. **Cron + push to chat surfaces.** Hermes pulls users back via daily digests delivered to Telegram. Waggle's signal bus + Launcher arc is the foundation, but we have no outbound digest path. **Action:** wire scheduled agents → email/Slack/Telegram delivery (the AI-OS arc opens the door; finish the loop).
3. **Open skill standard (`agentskills.io`).** Community-portable skills + the `awesome-hermes-agent` curated list. Waggle has skills internally but no public registry / marketplace front-door. **Action:** publish skill format + a public registry — this is the network-effect moat we keep deferring.
---
**Sources:**
- <https://github.com/NousResearch/hermes-agent> (v0.14.0, 170k stars, MIT)
- <https://hermes-agent.nousresearch.com/> (marketing copy, gateway list)
- <https://github.com/NousResearch/hermes-agent/issues/10488> (web-onboarding gap)
- <https://github.com/NousResearch/hermes-agent-self-evolution> (DSPy + GEPA companion)
- <https://github.com/0xNyk/awesome-hermes-agent> (community skill registry)
- <https://portal.nousresearch.com/manage-subscription> (Nous Portal hosted tier)
- TokenMix / innobu / Fastio / MindStudio / userorbit / DEV.to reviews (day-30 retention, addictive loop)