Hermes Agent
Hermes removes Collective Wisdom from core and proposes a smaller consent-gated plugin
A merged removal takes the organization skill marketplace out of Hermes core. A fresh proposal brings back a narrower instruction-sharing plugin, retaining explicit consent and hash verification while dropping proactive advice queues and multiple UI surfaces.
Hermes has changed direction on where shared organizational skills belong. PR #108507, merged September 11, removes the large Collective Wisdom implementation from core. The follow-up proposal #108678 brings the idea back as roughly a thousand lines in a bundled plugin directory, with no core edits. This is an architectural and product-scope reversal, not merely a smaller release number.
The facts
- The merged removal says Collective Wisdom belongs in a plugin or the MCP/connectors catalog rather than the core agent. - The replacement proposal limits packages to instructional files such as SKILL.md, a skill manifest and text under refs or assets. - It verifies individual blobs and the assembled package before writing an installation to disk. - Install, update, uninstall and share operations require a confirmation callback and refuse to mutate without it. - The proposed tools are hidden until entitlement is present, and unattended model-tool approval is intended to fail closed. - Proactive advice queues, weekly agent review, chat-card adapters and Desktop/dashboard panels are deliberately omitted from the smaller proposal.
Why it matters
Shared skills are executable influence even when they are not executable files. They can change which tools an agent chooses and what it considers normal operating procedure. Separating that exchange from core reduces the product surface that every installation must carry, while explicit consent makes distribution a user decision rather than an invisible learning side effect. Hash checks establish package integrity; they do not establish that the instructions are wise, appropriate or safe for a particular organization.
Current
Inspected on 2026-09-12. The Hermes Agent release baseline is v2026.9.11 published 2026-09-11T19:20:31Z. The main source was open (not merged) when captured. Proposals, merged mainline changes and packaged releases are different availability states; the linked release baseline does not imply that open proposals have shipped.
Evidence
Primary evidence: NousResearch/hermes-agent PR #108678 (https://github.com/NousResearch/hermes-agent/pull/108678). Supporting context: NousResearch/hermes-agent PR #108507 (https://github.com/NousResearch/hermes-agent/pull/108507); NousResearch/hermes-agent PR #94266 (https://github.com/NousResearch/hermes-agent/pull/94266). Sources were read directly through public pages or GitHub API bodies. Contributor test results are attributed claims, not independently reproduced results.
Operator take
Evaluate the reduced contract on its own terms. An organization considering it should review allowed content, who can share, who can install and how a revoked or updated skill affects active sessions. The author reports instruction-only rejection tests, byte-exact hash-vector checks and mocked HTTP end-to-end approval cases. The same source explicitly says a real Wisdom-enabled Gateway round trip was not exercised. That boundary is useful: the proposal can be assessed as a smaller implementation without claiming a proven organizational deployment or the return of the removed proactive product surfaces.
Caveat
The removal is merged; the smaller replacement was still open when captured. The older large implementation is cited to explain the reversal, not as a current capability list. A hash-valid instruction package can still contain poor advice, and the proposal's consent guarantees are implementation claims that have not been independently tested in this report.
The removal is merged; the smaller replacement was still open when captured. The older large implementation is cited to explain the reversal, not as a current capability list. A hash-valid instruction package can still contain poor advice, and the proposal's consent guarantees are implementation claims that have not been independently tested in this report.