Hermes Agent
by Nous Research · Agent · Text & code
promising
Fahd Mirza reuses it as a test harness for many models; little else is known
New /loop command re-runs a prompt on a recurring timer, re-reading current state each wake-up and reporting back
AccessOpen codePricefree and open source; Nous Portal Plus $20/monthRuns onWindows, macOS, LinuxLicenseMIT
Why we say so
Fahd Mirza ran it hands-on with Qwen4 Max, Hy4 flash, Maxwell V1.0 Flash, Apex 35B and Gemini 4 Neon, including screen control and sub-agents, and no problems were reported. The only negative is a report from @DanielLenton1 that it failed to memorize or generalize on Continual-ARC, which is a narrow test. One tester supplies nearly all the evidence, and no one has compared it with other agents.
Evidence
- Nov 19Used as the harness for Hy4 flash, including computer use of the screen with clicking and typing in real apps
- Sep 8Demo watched a simulated four-stage deploy; loop ended itself after five wake-ups, switching to file mtime checks when reads were deduplicated
- Sep 2Paired with Qwen 3.8 Flash Next, called the local agentic gold standard by the reviewer for speed
- Aug 21Commenter says Hermes currently has the upper hand over OpenClaw among people building multi-agent systems
- Aug 18/loop stops after a fixed count, a natural-language condition, a judge model check, or when the agent decides it is done
Show all 7
- Aug 18Rule of thumb: /loop for timer-driven watching, /goal for judge-driven fixing, cron for unattended work that outlives the terminal
- Aug 1With a local Qwen 27B via Ollama, it picked the Obsidian skill and found the vault unprompted for a plain note request