55fd0badecc6cd51...

🧵 how hermes agent gets better the longer it runs most ai agents start fresh every session. hermes doesn't. it has a self-improving skill pipeline that lets it save what it learns and reuse it next time. here's how it works: ◇ when does it trigger? the agent's system prompt tells it: after completing a complex task (5+ tool calls), fixing a tricky error, or discovering a non-trivial workflow — save the approach as a skill. this isn't optional metadata. it's hardcoded behavioral guidance injected into the system prompt whenever the skill management tool is loaded. ◇ what is a skill? a skill is a directory with a SKILL.md file — markdown with YAML frontmatter. it can include: - the main instructions (step-by-step approach) - reference docs (api details, cheat sheets) - templates (reusable configs) - scripts (automation code) skills follow the agentskills.io standard — portable, auditable, shareable. ◇ what happens under the hood? 1. the agent discovers a new workflow or solves a hard problem 2. it calls skill_manage(action='create') to write a SKILL.md to ~/.hermes/skills/ 3. the skill gets YAML frontmatter: name, description, version, tags 4. next session, it scans ~/.hermes/skills/ for all SKILL.md files 5. builds a structured index: metadata first (cheap), full content on demand 6. when a matching task comes up, it loads the skill and follows the saved approach ◇ the patch loop if a skill is outdated, incomplete, or wrong — the agent patches it immediately. not on the next session. not when asked. right now. the prompt says: "skills that aren't maintained become liabilities." ◇ progressive disclosure (token efficiency) skills use a three-tier loading system to save context window space: - tier 1: name + description (shown in skills list, ~minimal tokens) - tier 2: full SKILL.md content (loaded when relevant) - tier 3: linked files (loaded only when specific files are needed) this means the agent knows what tools it has without burning context on full instructions it doesn't need yet. ◇ the bundled → user pipeline on install, hermes ships ~25 bundled skills across domains (mlops, github, research, etc). these get synced to ~/.hermes/skills/ via a manifest that tracks hashes. if a bundled skill updates but the user hasn't modified their copy, it auto-updates. if the user customized it, their version is preserved. ◇ what this means in practice you use hermes for a week. it solves a tricky docker networking problem. it writes a skill about it. next week, when you hit the same problem, it doesn't rediscover the solution — it loads the skill and executes the saved approach. the agent literally gets better at its job over time. source: github.com/NousResearch/hermes-agent