Hermes Agent is INSANE...

TL;DR

Hermes Agent is a persistent, self-learning AI agent that can autonomously build projects, run benchmarks, and orchestrate other AI tools like Claude Code and Codex overnight.

Key Points

• 1.Hermes Agent grows smarter over time through persistent memory and auto-generated skills. Built by Noose Research, it learns your projects, iterates on solutions like a science experiment, and never forgets how to solve a problem — unlike stateless chatbots.
• 2.The creator built an entire AI benchmarking simulation called Grav (Gravity Well) using only AI agents. The simulation features four suns, physics-based ships with fuel limits, conservation of momentum, and a moving target circle — all coded without the creator touching more than 3 lines manually.
• 3.Models are benchmarked by iterating 20 times on ship-piloting code, then tested across 100 different seeds. Claude Opus 4.5 topped out at 276; Claude Sonnet 4.6 peaked around 78; Claude Sonnet 4.5 scored just 1 on its first run — illustrating massive performance gaps between models.
• 4.AI agents ran all benchmark simulations autonomously overnight from 2:17 a.m. to 5:32 a.m. Models tested include GPT 5.5 Pro, GPT 5.4, Grok 420, Deepseek V4 Pro, Gemini 3.1 Pro Preview, and multiple Anthropic models — all without human intervention.
• 5.Installing Hermes Agent on a Hostinger KVM2 VPS ($8.99/mo) requires just a few terminal commands. The recommended setup uses Ubuntu 24.04 LTS, 2 vCPU cores, 8GB RAM, and 100GB NVMe — with Noose Portal replacing the need for separate API keys for web search, image generation, and browser automation.
• 6.Noose Portal bundles web search, image generation, text-to-speech, and browser automation into one subscription. Unlike Open Router, it removes the need to configure individual API keys for services like Firecrawl or Browser Use — and currently offers Kimi K2.6 free for 24 hours via a partnership.
• 7.GPT 5.5 can be enabled inside Hermes Agent via OpenAI Codex OAuth login, and GPT Image 2.0 can be set as the image generation tool. Running 'hermes update' then selecting the provider and model is all that's required; no separate API key is needed for image generation at medium/low/high quality settings.
• 8.Hermes Agent orchestrated a live duel benchmark pitting GPT 5.5 High (via Codex) against Claude Opus 4.7 (via Claude Code) across 10 iterations. GPT 5.5 High won 7 of 10 rounds with a score of 68 vs. Claude's 43 — and Hermes automatically created a reusable skill called 'Gravo GPT Agent Loop' for future runs.
• 9.Running agents with safety confirmations disabled on isolated machines (VPS or old laptops) is the creator's preferred approach. He SSHes into remote servers, keeps API keys off shared systems, and uses tools like 1Password and Docker sandboxing to contain potential 'blast radius' if an agent causes damage.