本文由 AI 分析生成
建立時間： 2026-06-19 來源： https://x.com/Mnilax/status/2063697740526399833
Summary
A practitioner’s log of running Nous Research’s Hermes Agent (a self-hosted, persistent, scheduled agent) for 5 weeks on a $5 VPS with Claude as the underlying model. Provides 17 copy-paste prompts for standing jobs (morning briefs, repo watching, inbox triage, on-call diagnosis) and three lessons learned: vague schedules cause notification floods, hourly jobs need token budgets, and cheap local models fail multi-step tool calls.
記錄作者用 Nous Research 的 Hermes Agent（自架、持久化、可排程的代理）在 $5 VPS 上跑了 5 週，底層模型是 Claude。提供 17 個可直接貼上使用的常駐任務提示詞，並總結三個教訓：模糊的排程指令會造成通知洪流、每小時任務需要設定 token 預算、便宜的本地模型在多步驟工具呼叫上會失敗。
Key Points
- Hermes differs from session-based tools (Claude Code, ChatGPT) by being persistent, scheduled, and reachable via chat platforms — it keeps working when the session/tab is closed.
- Every working prompt to a standing agent needs three parts: a trigger (schedule/event), a body (what to do), and an explicit escalation rule (when to actually bother the human) — without the third part you get a firehose.
- Swapping the underlying model to a frontier model (Claude Opus) fixed multiple “broken” recipes at once; the failures were never the prompts, they were the model dropping tool calls mid-task.
- A serverless/hibernating backend (e.g., Daytona) keeps 24/7 idle cost near zero versus an always-on VPS.
- A good run can be promoted into a reusable named skill (“save this as a skill called X”), so the format doesn’t need re-explaining each time.
Insights
This is the same loop-engineering pattern (trigger → action → verify/escalate → memory-as-skill) showing up in a consumer/personal-automation context rather than a coding context — the verification gate here is just “human escalation rules” instead of tests, which is weaker and explains why vague schedules failed first. The fact that model quality (not prompt wording) fixed every broken recipe at once reinforces that loop reliability is bottlenecked by the model’s tool-calling robustness, not the prompt design.
Connections
- Loop engineering the 14-step roadmap from prompter to loop designer.
- Loops explained Claude, GPT, Mira and what actually works
- loop-engineering
- ai-agent
Raw Excerpt
A prompt to a chat window is a question. A prompt to a persistent agent is a job description: it needs a trigger (a schedule or an event), a body (what to do), and an escalation rule (when to bother you). Drop any of the three and the prompt either never fires, does the wrong thing, or buries you in noise.