Hermes v2026.7.1 Update: Hermes Agent v0.18.0 (2026.7.1) — The Judgment Release

One-sentence summary: Hermes Agent v0.18.0 is a massive clean-sweep release that fixes every single critical and high-priority bug in the entire project, while also making the AI dramatically smarter at thinking and knowing when it’s truly done with a task.

🎯 The Big One: Zero Critical Bugs Left

Imagine your car has a check-engine light that’s been on for months. Now imagine a team of mechanics spends 12 days straight fixing every warning light, not just the obvious ones. That’s what this release did.

The Hermes team closed ~692 highest-priority issues and pull requests in just under two weeks. That includes 3 critical (P0) bugs and 493 high-priority (P1) issues, plus 196 high-priority PRs. The result? Zero open P0s and zero open P1s across the entire repository. They even fixed a nasty bug where compressed data could corrupt sibling processes (issue #56391) — a classic “two things talking at once” problem, fixed with an all-nighter right before release.

Special shoutout to @kshitijk4poor, who helped burn through the backlog like a human flamethrower.

✨ What’s New: Smarter Thinking, Clearer Done

1. Mixture-of-Agents is now a model you pick (not a mode you toggle)

Previously, using multiple AI models together was like manually assembling a committee every time. Now, it’s like having a pre-formed “dream team” on speed dial. You can create a named ensemble (e.g., “my-council”) and select it in any model picker — CLI, desktop, or gateway — just like you’d pick Claude or GPT. Your prompt automatically goes to the whole council, you see each member’s reasoning, and the final answer streams live. It’s like asking three experts to debate before giving you a verdict — but with one click.

2. The agent verifies its own work against evidence

Before, Hermes might finish a task and say “done!” based on vibes. Now, it checks its work against actual evidence — like a student re-reading the textbook before saying “I know this.” This means fewer hallucinated completions and more trustworthy results.

3. /goal gets completion contracts

You can now define exactly what “done” means for a goal. Think of it as a checklist that the agent must tick off before it’s allowed to stop. No more “close enough” — it either meets the contract or keeps working.

4. /learn and /journey make self-improvement visible

The agent can now show you what it’s learning and how it’s evolving over time. It’s like having a fitness tracker for your AI’s brain — you can see the progress, not just trust it.

5. Gateway scales to zero, desktop gets coding projects

The gateway (the server that connects everything) can now scale down to zero when idle — saving money when nobody’s using it. The desktop app now has first-class coding projects (like a mini-IDE) and a playable memory graph — you can literally see how your agent’s memories connect.

6. Background subagent fan-out

Subagents can now work in the background, fanning out to do multiple tasks at once. It’s like hiring a team of assistants who each go off and handle a piece, then report back — without blocking you.

🔧 Fixes & Improvements

Beyond the P0/P1 sweep, this release includes ~1,720 commits, 998 merged PRs, and fixes for 949 issues. That’s a lot of polish — from cron reliability (scheduled tasks actually firing on time) to credential-exfiltration hardening (keeping your secrets secret).

📌 What This Means for You

If you’ve been holding off on Hermes because of a nagging bug, now’s the time to jump in. The foundation is rock-solid. And the new “thinking” features mean you’ll spend less time double-checking your agent’s work — it’s doing that itself now. For power users, the MoA presets and completion contracts are game-changers for complex, multi-step projects.

Bottom line: This is the release where Hermes grew up — from “promising framework” to “dependable, self-aware workhorse.”

📎 Official release notes: https://github.com/NousResearch/hermes-agent/releases