Hermes Agent Update v0.18 is HUGE! (Judgment Release)

Hermes Agent v0.18 "Judgment" closes ~700 P0/P1 issues in a 12-day sprint and ships first-class Mixture-of-Agents (MOA) presets, evidence-based task verification, and Vertex AI/Gemini auth.

Stop Using a Single Model — Configure /MOA First

MOA is now a dashboard-level preset, not a hack. Run /MOA on your first prompt for multi-step work so an aggregator model (e.g. Kimi 2.7 for cheap coding calls) synthesizes outputs from your full council of reference models. Reference models don't get tool access — the aggregator handles that. The default aggregator routes through OpenRouter, so if you have no OpenRouter keys configured, /MOA silently fails. Skip MOA for trivial one-step prompts.

Treat Completion as Evidence, Not Claims

The /goal command now supports completion contracts: pass/fail defined by concrete artifacts — test logs, exit codes, generated files. Hermes records a verification audit trail, so you can ask "did you actually run the test?" and get proof instead of a vibe answer. This is the difference the creator calls between "an agent that claims and an agent that proves."

Self-Improvement Loop Got Cheaper

The post-turn learning loop no longer replays the full conversation every turn — an auxiliary model digests the context, not the full transcript, then adapts cadence. Token burn drops significantly. Creator's caveat: if you're the type who wants raw full-transcript review, you'll feel the loss.

/journey Memory Is Useless If You Reinstalled Your OS

/journey exposes a timeline of skills, memories, corrections, and preferences you can audit and prune. Creator reset his server on June 11 and saw almost nothing — the feature only pays off after ~3 months of accumulated agent history. VPS re-installers will get limited value.

Vertex AI + Security Hardening You Shouldn't Skip

Google Vertex AI is now a first-class provider for Gemini — auto-mints short-lived OAuth2 tokens from your GCP service account, no static API key in .env. Security patches target MCP config-persistence attacks, cron base-URL exfiltration, secret prefix leakage, Slack app-token redaction, and a raised AIO HTTP floor for CVEs. If you've pasted real credentials into chat, this section is the most important one in the release notes.

Operational Wins

Gateway now scales to zero when idle and coordinates graceful drain before restarts — lower idle cost, safer deploys on VPS or self-hosted setups. Desktop app now has per-profile project lists with git worktree management, review panes, and persistent project context — "cursor light with Hermes's skill system and memory behind it." /prompt opens your local editor (Vim/VS Code) for long structured markdown prompts — local-machine only, not usable on VPS.

Background Sub-Agent Fan-Out

"Fan out sub agents" is now a recognized trigger phrase that activates your cloud code orchestrator's sub-agent mode, mirroring the Claude Code pattern.