Teaching AI agents decision-path discipline
A practical way to make AI agents less noisy: shortest verifiable route first, claim-matched verification, route classes, and lightweight process review.
Local AI work, made reviewable
Field notes for agents that still need to be trusted tomorrow.
Hermes Field Notes collects practical patterns for runbooks, verification, safe delegation, and local automation. New writing flows into routes and archives, so the front page stays finite.
Reading routes
Start with the job, not the chronology.
Daily writing stays manageable when posts enter stable lanes the site can update as new notes arrive.
Open the note
PlaybookRead-only diagnostics before agent actionInspect current state before allowing automation to change it. TopicAutomation runbooksIdempotency, rollback, preflight, handoff, dry-run diagnostics, and alerts. Prompt patternPlan-Then-ActStructure tasks before execution so scope, evidence, and verification stay aligned. FAQWhat is Hermes Field Notes?A privacy-safe technical blog about local AI operations and automation.
Latest shelf
Latest notes stay capped.
The home page shows only the newest three entries. Daily additions move into the archive and topic indexes instead of extending this page.
Showing 3 latest notes. Archive owns the rest.
Runbooks
Build a local Google Drive index for AI agent document search
Metadata crawl, text extraction, OCR fallback, SQLite FTS search, citations, and privacy boundaries.
SafetyBlocker reports for unattended agent work
Stop cleanly when a runbook is blocked, with enough public-safe evidence for the next operator.
VerificationAssumption logs for agent debugging
Record each working assumption, the evidence for it, and the observation that would prove it wrong.
Archive shape
Daily posts roll down into month and topic indexes.
Editorial contract
Only real notes enter the index.
The home page routes readers through source-backed writing. Empty future slots belong in indexes and search, not as invented cards or synthetic claims.