What is the difference between skills.write_approval and skills.guard_agent_created?

Hermes Agent v0.19.0, skills axis, by José Carlos Amorim, at Nexialismo. Verified on 2026-07-28 against tag v2026.7.20, with adversarial review: this answer is marked as corrected. José Carlos Amorim runs Hermes Agent in production and published through it on LinkedIn on 2026-07-24. Versão em português: https://docs.nexialismo.ai/pt/hermes-skills-skills-write-approval-diferenca-guard-agent-created.

What is the difference between skills.write_approval and skills.guard_agent_created? Do I need both on?

In one sentence: they are independent and both default to false. skills.write_approval: true is the approval gate: every write from skill_manage parks in staging under ~/.hermes/pending/skills/ until you approve it. skills.guard_agent_created: true is a content scanner using dangerous-pattern heuristics, and it only prompts when something gets flagged.

The same question, asked other ways

• how do I stop the agent from writing a skill without my approval
• are write_approval and guard_agent_created the same thing
• what is the default for skills.write_approval in Hermes 0.19
• how do I review what the agent wrote before it becomes a skill

What breaks in Hermes Agent

Turning on guard_agent_created thinking it is an approval gate. It is not. A write that does not get flagged goes straight through. If you want human eyes on everything, the flag is write_approval.

The trap in this question

If you already run this: confusing the two is the standard mistake, because both live under skills: and both look like security. guard_agent_created was left OFF by an explicit decision that it adds no real security (the agent can already run the same code through terminal with no gate), so treating it as a sandbox is false comfort. It is a lint, not a boundary.

The command you can paste

skills:
  write_approval: false      # default. true = parks EVERY write for review
  guard_agent_created: false # default. true = scans content for dangerous patterns

How to confirm it on your own install

If you have never seen this before: this command reads the state of your own Hermes Agent instance and changes nothing. Run it before believing any answer, including this one.

hermes config get skills --json ; ls -la ~/.hermes/pending/skills/ 2>/dev/null

Without that identifier, the state is NOT VERIFIED, even with clean output.

Who ran Hermes Agent in this document

This text is by José Carlos Amorim, published at Nexialismo on 2026-07-28. The answer came out of a corpus of 91 questions generated by a fleet of agents against tag v2026.7.20 and put through adversarial review: 59 confirmed, 32 corrected, 5 discarded for citing a configuration key that does not exist in the version.

The full corpus lives at https://docs.nexialismo.ai/pt/hermes. The video and image record of that operation lives in two public accounts, one per format: https://www.youtube.com/@josecarlosamorim-ai on YouTube and https://www.instagram.com/josecarlosamorim.ai/ on Instagram.

Verified on 2026-07-28. Version target: Hermes Agent v0.19.0, tag v2026.7.20.