As my work on the llm-knowledge-base has grown, I’ve been shifting toward using Hermes (the agent) as the primary actor for working on the knowledge base, while still using Grok for higher-level reasoning when needed.

Core Idea

I’m treating Hermes as the agent and giving it the ability to use different models and tools depending on the task. This setup supports the longer-term goal of gradually increasing Hermes’ autonomy.

Current Division of Labor

Task TypePrimary ActorModel / ToolNotes
High-level framing & philosophyHermes (with Grok)GrokHermes can delegate complex reasoning to Grok
Writing new synthesis pagesHermesGrok or Hermes 3Depends on complexity
Auditing and improving existing hubsHermesHermes 3 (local)Primary execution work
Integrating new conceptsHermesGrok (planning) + Hermes 3Hybrid within the same workflow
Link hygiene, taxonomy, consistency workHermesHermes 3 (local)Fast, local, low-risk tasks
Running experiments across the systemHermesHermes 3 (local)Full local repo access
Deep diagnostic or philosophical workHermesGrokWhen stronger reasoning is required

Why This Approach

  • Hermes is the persistent agent. It can maintain long-term context about the knowledge base and gradually take on more independent work.
  • Grok is used as a high-quality reasoning tool that Hermes (or I) can call when needed.
  • Hermes 3 (local) is used for most execution work and anything involving private material. File editing tools can be invoked by Hermes when needed.

This structure supports the direction of starting with propose-and-review workflows and gradually increasing Hermes’ autonomy over time.

Current Setup (as of May 2026)

  • Hermes (the agent tool) is the main interface.
  • It can call Grok via API when deeper reasoning is required.
  • It can use the local hermes3:8b model via Ollama for most execution work.
  • File editing tools are available when direct changes are needed.
  • All work is currently in a controlled “propose and review” mode, with the intention of increasing Hermes’ independence as the system matures.

Open Questions

  • How much persistent memory and context does Hermes need before it can reliably handle larger autonomous tasks?
  • How should Hermes decide when to use Grok versus the local model for a given subtask?

Current Guardrails (as of May 2026)

Strict read/write boundaries are now defined in AGENTS.md → Hermes Access Boundaries:

  • Hermes may read wiki/ freely and read specific private files only when explicitly directed by the user.
  • All durable content changes target wiki/.
  • Raw agent activity (transcripts, proposed diffs, logs) is written to raw/sessions/.
  • After substantial work, Hermes must create a structured session summary using the template in raw/sessions/templates/session-summary.md and present it for review (see AGENTS.md “Session Summary Habit”).
  • Level 2 work requires explicit human approval on every wiki/ edit.
  • Private material handling remains human-gated (no autonomous crawling of raw/private/).

These boundaries are the foundation for safely increasing autonomy.