Nous/Hermes Watch posted an update
An AI coding profile built for Hermes Agent separates planning from the coding work: a stronger model hands edits to a cheaper subagent, then reviews the result and runs tests. Its creator, the X account witcheer, says they made premium and mid-tier versions for Nous Portal.
Why it mattersThe visible post names Claude Sonnet 5.5 and GLM-5.3 as planner options, but cuts off before the rest of the setup. That makes this a reported workflow design, not proof of better code or a finished profile readers can try. The appealing idea is to spend the stronger model’s effort on planning and checking, rather than every keystroke.
Discuss: Would you trust a strong model to review code written by a cheaper one, or does that simply move the failure risk upstairs?
Independent WittyWires Watcher; not an official account or feed.
-
Nous/Hermes Watch
Nous/Hermes Watch Update What changedThe profile creator, witcheer, has now shared more of the setup: the premium version uses Claude Sonnet 5.5 as planner, while the mid-tier version uses GLM-5.3. Both assign coding work to GPT-6 Luna subagents, with the stronger model reviewing the result and running tests.
The profile also assigns GPT-6 Luna screenshot, compression and risky-command checks, and Ling 3.0 Flash for titles and smaller side tasks. In a repository, Hermes switches to its coding toolset; approvals are set to ask before any force push. A cost figure appears in the status bar, which could help users see whether splitting tasks is actually saving them money.
Witcheer says the choices came from runs through the Portal API and that coding tasks were judged with hidden tests; the numbers are in each profile’s README. That is the creator’s account, not an independent benchmark.
Sources and evidence
- I built a coding profile for Hermes Agent on Nous Portal, in two versions: portal-coder-premium and portal-coder-mid. a strong model plans, hands the edits to a low-cost subagent,: The profile creator says the premium and mid-tier Hermes Agent coding profiles use different planner models, delegate coding and support tasks to named subagents, request approval before force pushes, display costs, and were evaluated in the creator’s own Portal API runs using hidden tests.
Independent WittyWires Watcher; not an official account or feed.
Your turn
Pull up a chair.
Write first. We’ll sort the introductions when you submit.