<- all tokdocs

Hermes Agent Self-Improving Skills: Background Review vs the Curator

By , founder of Mavric and Apso

View on TikTok ->

Hermes Agent, the open-source agent from Nous Research, does write its own skills, and a TikTok from NetworkChuck shows it happening after he connects the agent to his home lab through Twingate. His video description credits the Curator feature for it. The Hermes repository shows the skill was created by a different subsystem, the background review, while the Curator only maintains and prunes skills on a seven-day cycle.

What happens in the video

Chuck runs Hermes on a Hostinger VPS, personified as Ron Weasley, using OpenAI's GPT-5.5 as the model. He gives it a Twingate service key, and the agent sets up a headless Twingate client and reaches his home network, named Hogwarts, which runs UniFi, MikroTik, Ceph and Home Assistant. He then types "Search the network, see what you can find," and the terminal prints:

💾 Self-improvement review: skill "twingate-client-operations" created

He says "I didn't tell him to do this," and closes with "That's the power of Hermes. It makes its own skills."

Does Hermes Agent create its own skills?

Yes. The repository tagline is "The agent that grows with you," and the homepage says "It learns your projects, auto-generates skills, and never forgets how it solved a problem."

The skills documentation explains the mechanism. Skills live in ~/.hermes/skills/, "the primary directory and source of truth." The agent writes them with a tool called skill_manage, which supports create, patch, delete, write_file and remove_file. It creates one when it "worked out a multi-step workflow worth repeating," "hit errors or dead ends and found the working path," or when "the user corrected its approach." A headless Twingate setup fits the first trigger.

The skills.write_approval setting defaults to false, which is why the agent did not ask before saving the skill.

Background review vs Curator: which one wrote the skill?

The line on screen comes from the background review. In the repo's locales/en.yaml, the "Self-improvement review" template sits under a comment naming agent/background_review.py as its source.

cli-config.yaml.example describes the background review as a "Post-turn memory/skill self-improvement review fork. Runs after a turn when the nudge intervals fire; writes skills/memories in a daemon thread." It is enabled by default.

The Curator documentation describes a different job: "a background maintenance pass for agent-created skills. It tracks how often each skill is viewed, used, and patched, moves long-unused skills through active → stale → archived states." It adds that the Curator "exists so that skills created via the self-improvement loop don't pile up forever."

So the video description's "removing the ones that don't work" describes the Curator. "Creating new ones automatically" describes the background review.

Curator default settings

The Curator's defaults also rule it out as the cause of what happened on camera:

curator:
  enabled: true
  interval_hours: 168          # 7 days
  min_idle_hours: 2
  stale_after_days: 14
  archive_after_days: 30
  consolidate: false
  prune_builtins: false

It runs every seven days, only after the profile has been idle for two hours, and skips while a turn is running. On a fresh install, "the curator does not run immediately" and defers its first pass by one full interval. Its consolidation pass, which merges overlapping skills, is off by default "because it costs aux-model tokens on every run and makes broad structural changes to your library." The skills doc lists three possible writers: a foreground turn, the background review, and the Curator's consolidation pass.

Did the agent write the skill unprompted?

Mostly. The prompt Chuck typed says nothing about skills. One turn earlier, though, the agent had asked "Want me to save this Twingate headless setup process as a re[usable spellbook]" and got no answer. The background review then wrote the skill at the next turn boundary.

The video never opens the skill file, lists skills, or runs the skill again, so whether it works is not shown. The optional skills catalog has no Twingate skill, which supports it being newly written.

Is Twingate a VPN?

The video calls it "a VPN." Twingate describes "Identity-based access for users, services, and AI agents that deploys in minutes, scales to every resource, and finally lets you retire your VPN." It markets itself as Zero Trust access that replaces a VPN. The agent's output in the video shows resource-scoped routing on an sdwan0 interface to 10.77.14.0/24, rather than one tunnel for everything.

Which model does the demo use?

The status bar reads "gpt 5.5," which is OpenAI's GPT-5.5, released 23 April 2026. Hermes Agent is model-agnostic and supports Nous Portal, OpenRouter, OpenAI and custom endpoints. The agent shares a name with Nous Research's Hermes model family, but this demo does not use those models.

Hermes Agent is MIT-licensed. On 2026-10-26 the repository showed 257,220 stars, and the latest release was v0.22.5, published on 2026-10-23.

Bottom line

Hermes Agent writes reusable skills on its own through the background review, which runs after a turn and is on by default. The Curator is a weekly cleanup pass that archives unused skills and does not create them. Check the skill file before trusting it, because the video does not show it working.

Sources

Published October 23, 2026. Based on a TikTok by @networkchuck. Full video notes (markdown)