Hermes Agent
Hermes proposes task-level reasoning controls alongside desktop model selection
Hermes Desktop would let users set reasoning effort separately for auxiliary tasks, matching the runtime’s existing configuration keys. The proposal distinguishes leaving an override alone, clearing it to inherit and setting a validated level; related CLI work carries effort with the model pick and its scope.
A September 14 Hermes Desktop proposal places a reasoning-effort selector beside each auxiliary task’s provider and model choice. The runtime already understands task-specific effort, but the source describes a GUI that previously carried only provider and model in its draft and save request. The proposed change makes the existing control visible instead of requiring a configuration edit.
The related CLI proposal broadens the same idea across model pickers: a selection can carry its reasoning effort as part of one scoped request. Session, global and one-turn choices remain different operations. The Desktop work is narrower, focusing on auxiliary task settings and the API contract that persists them.
The facts
- The Desktop setting writes auxiliary.<task>.reasoning_effort, the same task-level key used by the runtime. - Omitting reasoning_effort leaves it unchanged; null clears the override to inherit; a recognized level sets it. - An unknown level is rejected with HTTP 400 instead of being silently dropped. - Reset all to main removes task-specific overrides while the task row shows an effort label when one is set. - The related CLI proposal adds a reasoning step after model selection and keeps one-turn, session and global scope distinct. - The Desktop author reports a real isolated backend and mounted settings page with the Electron bridge stubbed, plus 27 model-settings and 31 localization tests.
Why it matters
Choosing a cheaper model for a background task and choosing how much reasoning that task requests are separate decisions. A visible per-task control lets operators express that distinction without accidentally altering every conversation. The most useful part is not the size of the effort ladder but the save semantics: a request that omits an option should not erase it, and a request to inherit should not merely return success while keeping stale state.
Current
Inspected on 2026-09-14. The Hermes Agent stable-channel baseline is v2026.9.11 published 2026-09-11T19:20:31Z. The primary source was open (not merged) when captured. Mainline merges, open proposals and packaged releases are distinct availability states; the release baseline does not establish that a proposal has shipped.
Evidence
Primary evidence: NousResearch/hermes-agent PR #110435 (https://github.com/NousResearch/hermes-agent/pull/110435). Supporting context: NousResearch/hermes-agent PR #110346 (https://github.com/NousResearch/hermes-agent/pull/110346); Nous Research — Hermes product documentation (https://hermes-agent.nousresearch.com/docs/). Public source bodies and linked documentation were inspected. Contributor test results remain attributed reports, not independently reproduced experiments.
Operator take
After changing one auxiliary task, read it back and verify sibling tasks and provider/model selections are untouched. Test inheritance by clearing the override, not by substituting the main model’s current level as a fixed value. Provider routes differ in the effort controls they support; treat a saved setting as requested configuration, then verify the actual route honors it before claiming lower cost, higher quality or a particular latency change.
Caveat
The Desktop PR was open when inspected, and the companion CLI proposal is supporting direction rather than a blanket availability claim. The browser test used an isolated server with a stubbed Electron REST bridge, not a packaged desktop acceptance run. No comparative model-quality, latency or billing experiment was performed for this article.
The Desktop PR was open when inspected, and the companion CLI proposal is supporting direction rather than a blanket availability claim. The browser test used an isolated server with a stubbed Electron REST bridge, not a packaged desktop acceptance run. No comparative model-quality, latency or billing experiment was performed for this article.