Project 2 min read
Hermes: Custom AI Agent Infrastructure
Open-source AI agent adopted, extended, and production-hardened for daily operations.
I run an open-source AI agent as daily infrastructure. The framework is Hermes Agent by Nous Research — I adopted it, configured it deeply, extended it with custom skills and MCP optimization, and wired it into every layer of my work.
I chose it after repeated failures with another agent framework that couldn’t hold up under real use. The difference: Hermes is open-core, extensible at the edges, and runs fully local. What started as a configuration exercise became a daily coordination layer.
What It’s Actually Done #
- Coordinated a full graduation project cycle end-to-end — generated documents, interviewed teammates, ran code, produced charts, conducted research, kept the whole thing on track
- Ran cybersecurity crisis analysis and hacker infrastructure mapping during a 161K-follower account recovery
- Compressed thesis production — citation synthesis across sources plus document generation — from a manual 1–2 week job to one day
- Turned dead time into work time: website edits, publishing, research — all runnable from a phone mid-commute
- Automated bulk operations that used to fight page-load times and tab-switching — ActiveCampaign batch labelling across thousands of contacts, bulk patching, bulk editing, file renaming
- Handles image generation and Anki card production, including voice modulation
- Runs a daily morning brief pulling trending news for niche and global awareness
Architecture #
Foundation — Hermes Agent v0.18.0 (Nous Research) #
- Open-source agent core: runs locally, extensible via skills, plugins, and MCP servers
- Multi-platform gateway: Telegram, Discord, WhatsApp, Signal, and ~20 other chat platforms
- Built-in tools: terminal, file I/O, web search, browser, code execution, task delegation
- Conversation loop with prompt caching, context compression, and memory
Custom Integration — My Additions #
Pruned from 11 MCP servers down to 4, converted the rest into on-demand skills. Result: context window usage dropped from ~16% to ~4–5% on a 1M-token model — roughly a 400% reduction in overhead.
Convert always-on MCP servers into on-demand skills when you only need them occasionally — the context window savings compound across every turn of a long session.
- Multi-agent A2A orchestration: configured Hermes as the coordination layer that dispatches coding tasks to Claude Code, Cursor, Antigravity, and Codex through an agent-to-agent protocol — each tool used for what it’s best at
- Hybrid memory architecture: Claude side runs claude-mem; Hermes runs Hindsight with built-in memory fallback, hard routing rules enforcing the path, and a daily cron job that consolidates and cleans memory automatically
- ~20 custom skills: authored for brand-specific operations — ActiveCampaign email automation, B2B sourcing pipeline, FIH pilot workflows, audio processing, Anki production, and more
- Production operations: 3,000+ successful cron executions, daily morning briefs, real-time alerting, and ongoing iteration