MP mikepatraw.com

Hermes Agent System / Case Study

No-LLM YouTube Trigger and Cost-Control Pattern

Replacing idle polling with lightweight event routing for simple automation triggers.

Draft public-safe Public-safe portfolio draft

Executive Summary

Simple structured events should not wake a model on a schedule. This case study shows how the YouTube learning workflow moved toward cheap event detection and reserved LLM reasoning for actual analysis.

Problem

What needed solving

Simple events like bare YouTube URLs should not require scheduled LLM polling. Polling wastes tokens, adds latency, and creates avoidable operational noise.

Context

Why this mattered

The YouTube learning workflow originally risked being treated as a recurring polling problem. The better pattern was a no-token trigger: route obvious URL events directly into the learning pipeline when chat/event infrastructure sees them.

Constraints

Operating boundaries

  • Do not wake an LLM on a schedule just to detect simple regex-matched events.
  • Keep YouTube processing inside normal chat/event routing when possible.
  • Preserve the transcript-first and Gemini-minimal fallback cost controls.
  • Avoid storing browser cookies or sensitive extraction credentials in plain text.

Build

What I built

  • A documented no-LLM Discord YouTube trigger pattern.
  • A lightweight URL detection helper and regression tests.
  • Skill guidance that prevents reintroducing LLM polling for bare YouTube links.

Architecture

Workflow

  • Trigger: chat message arrives containing a YouTube URL.
  • Detection: cheap regex/event hook identifies the URL without model inference.
  • Execution: YouTube learning pipeline runs transcript-first, then fallback video ingestion only if needed.
  • Output: ledger/demo artifacts and a concise report back to the originating thread.

Security / Operations

Key decisions

  • Use event routing for obvious structured triggers and reserve LLM reasoning for analysis after ingestion.
  • Document the failure mode directly in the skill so future automation does not drift back into polling.
  • Treat bot-detection and transcript failures literally: quote exact errors and fall back deliberately.

Impact

What changed

  • Reduced unnecessary token spend for idle detection.
  • Made the automation cleaner and easier to defend as a portfolio-quality PA system.
  • Separated trigger mechanics from learning analysis, which improves maintainability.

Evidence

Source trail

gemini-minimal cron · trigger

Hermes Agent Full Tutorial for Beginners | Setup Guide

crontrigger
View source
gemini-minimal trigger · event

Hermes Agent Just Got 10x BETTER (Curator Explained)

triggerevent
View source
gemini-minimal url · youtube

Hermes Agent: The New OpenClaw?

urlyoutube
View source

Next

What I’d improve

  • Move the regex hook deeper into gateway routing if normal chat instantiation misses bare URLs.
  • Add metrics for avoided polling/model calls.
  • Keep this as a subsection under the larger Hermes Agent System story.

Public-safe notes

Publishing boundary

  • The published case study avoids private Discord channel details and bot tokens.
  • The key public lesson is architectural: event routing for structured triggers, model reasoning for analysis.
  • Fallback behavior is documented so failures stay observable instead of being explained away.