Hermes Agent released v2026.4.30 on Apr 30, with Curator as the headline feature. Hermes describes it as the release where “Hermes Agent now maintains itself”: an autonomous background Curator grades the skill library, consolidates related skills, prunes dead ones, archives skills with consolidated-vs-pruned classification, and writes per-run reports.
This is publishable for Open Orchestrators because it is not just another integration or provider update. Hermes already treats memory, skills, schedules, subagents, and gateways as part of the agent runtime. Curator adds lifecycle management around the skills that accumulate as the agent learns from repeated work.
The core shift is operational. A long-running agent can create or update reusable skills, but without curation the skill library can become a hidden failure mode: duplicated instructions, stale workflows, narrow one-off skills, and old context competing for attention. Curator turns that into a first-class maintenance loop with usage-aware status, scheduled review, archive/restore behavior, pinning, and guardrails that protect bundled or hub-installed skills from mutation.
The release notes also frame Curator as part of a broader self-improvement upgrade. Hermes says the background review fork is now rubric-based, biased toward updating the skill the agent just loaded, able to handle references/ and templates/ sub-files, restricted to memory and skills toolsets, and better aligned with the parent’s live runtime configuration.
For Open Orchestrators readers, the category signal is clear: persistent agents need garbage collection, not only memory. If skills are the reusable operating layer for an agent, then skill lifecycle management becomes governance infrastructure. Curator gives Hermes a built-in answer for how a self-improving agent avoids turning its own learned behavior into a cluttered prompt closet.
For the Agent Analytics angle, the useful bridge is downstream of the agent work. Hermes can maintain the skills and scheduled loops that ship changes, while Agent Analytics measures the user-facing outcomes those changes create: visits, sources, onboarding completion, signup, activation, retention, and experiment movement. Curator keeps the agent’s operating playbook cleaner; Agent Analytics checks whether that playbook is improving the product surface.
Source confidence: High. The feature is documented in the official v0.12.0 release notes, Teknium announced Curator publicly on Apr 30, and merged Hermes PRs show the background skill maintenance, archive classification, and skill-usage status work landing before the release.
Evidence:
- Hermes Agent v0.12.0 GitHub release
- Teknium X announcement: Introducing Hermes Curator
- Hermes Curator documentation
- Hermes PR #17277: background skill maintenance
- Hermes PR #17941: consolidated vs pruned archive classification
- Hermes PR #18033: most-used and least-used skills in curator status
- Hermes Agent GitHub repository
Explicit non-claims:
- This update does not claim Hermes invented reusable agent skills or agent memory.
- This update does not claim Curator can prove that a skill is semantically correct, safe, or optimal for every project.
- This update does not claim Agent Analytics measures Hermes memory, traces, logs, or skill internals.