Hermes 0.18 "The Judgment Release": AI Agents Learn to Prove Their Work Is Done

Nous Research has released Hermes Agent v0.18.0 ("The Judgment Release"), clearing 700 P0/P1 issues in 12 days. The update introduces completion contracts for self-verification, Mixture-of-Agents as a first-class citizen, and a suite of new features including /learn, /journey, and desktop Projects.

💡 What You Will Learn

Nous Research has released Hermes Agent v0.18.0 ("The Judgment Release"), clearing 700 P0/P1 issues in 12 days. The update introduces completion contracts for self-verification, Mixture-of-Agents as a

Nous Research has released Hermes Agent v0.18.0 (The Judgment Release), named after the team's aggressive 12-day sprint that cleared all 700 P0 and P1 issues — 3 issues + 8 PRs for P0, and 493 issues + 188 PRs for P1. The release encompasses 1,720 commits, 998 merged PRs, 250,000 lines of new code, 949 closed issues, and contributions from 381 developers.

Key upgrades include:

Mixture-of-Agents (MoA) as a First-Class Citizen — Previously an experimental mode, MoA presets now appear as selectable models in the model picker. Each reference model's reasoning is displayed in its own tab, and the final aggregated answer streams in real-time. This allows users to see not just the committee's conclusion but how each "member" voted and why.

Completion Contracts — The most underrated upgrade. Agents can now be given explicit completion criteria via /goal. Hermes actually runs tests and checks evidence rather than relying on the model's own judgment that "it's done." A pre_verify hook allows custom validation pipelines. This shifts agents from "I think I fixed it" to "tests passed, code review done, CI green, here's the evidence."

/learn and /journey Commands — /learn lets users teach Hermes a directory, workflow, or URL, which is automatically distilled into a reusable Skill. /journey opens a learning timeline in CLI and desktop, showing accumulated memories and skills, with an interactive radial timeline visual.

Desktop as a Programming Cockpit — Native Projects feature with sidebar codebase management, review panels, git worktree management, multi-terminal panels. Persistent window state, auto-TTS for reasoning, and PR-style file diffs embedded in chat.

Background Fan-Out — delegate_task now spawns multiple sub-agents that run in the background without blocking the chat, aggregating results when all finish.

Google Vertex AI as a first-class provider with automatic token refresh.

Gateway Scale-to-Zero — Auto-sleeps when idle, wakes on activity, enabling true pay-per-use for cloud-hosted Hermes.

Security — MCP config persistence hardening, cron base_url injection protection, Slack xapp- token auto-redaction.

Personal Take: The most "judgment-like" feature isn't the bug cleanup — it's the self-verification mechanism. An AI agent becomes a trustworthy worker not when it gets smarter, but when it can say "I'm done, here's the evidence." This release shows Nous Research is seriously thinking about how to make AI agents reliably accountable.