Original release notes
Hermes Agent v0.18.0 (v2026.7.1)
Release Date: July 1, 2026
Since v0.17.0: ~1,720 commits 998 merged PRs 2,215 files changed ~251,000 insertions ~41,000 deletions 949 issues closed 370+ community contributors
The Judgment Release. Over the last week and a half the team put nearly all of its effort into one goal: resolve every P0 and P1 issue and PR in the entire Hermes Agent repo - and as of this release, 100% of them are closed. Zero open P0s. Zero open P1s. That's ~700 highest-priority items cleared as part of ~1,950 total issues and PRs closed this window. We intend to keep P0/P1 at zero from here on. On top of that clean-sweep, v0.18.0 is about how well Hermes thinks and how it knows when its work is actually done. Mixture-of-Agents became a first-class citizen - named ensembles of models you can pick like any other model, with every reference model's reasoning shown to you and the aggregator's answer streamed live. The agent learned to verify its own work against evidence instead of vibes,
/goalgained completion contracts, and/learn+/journeyturned self-improvement into something you can see and steer. Underneath, the gateway became genuinely deployable-at-scale (scale-to-zero, drain coordination), the desktop grew first-class coding projects and a playable memory graph, and subagents can now fan out in the background.
The P0/P1 Clean Sweep - 100% resolved
This is the release headline. For a week and a half the team hammered the priority backlog day and night, and every single P0 and P1 across the whole repo is now closed:
Priority Issues closed PRs merged
P0 (critical) 3 8
P1 (high) 493 188
Total 496 196
That's ~692 highest-priority items resolved in twelve days - and at the moment the sweep completed, the open P0/P1 count hit 0 across the entire repo. The final cluster to fall was the interrupt-protected-compression sibling-fork bug (issue #56391) and its fix (#56416), closed on an all-nighter right before this release cut.
Special shoutout to @kshitijk4poor, who burned through the priority backlog day and night alongside the core team - the cron reliability wave, the compression-fork fix, the credential-exfil hardening, and a huge share of the P1 closures are his.
We're keeping P0/P1 at 0 from here forward.
Highlights
Mixture-of-Agents is now a first-class model you can pick - MoA used to be a mode you toggled; now every named MoA preset shows up as a selectable model under a moa provider, right alongside Claude, GPT, and Grok in every model picker (CLI, TUI, desktop, gateway). Pick "my-council" the same way you'd pick any model, and Hermes routes your prompt through that ensemble automatically. An ensemble of frontier models deliberating on your hardest questions is now one selection away, on every surface. (#46081, #53548, #53561 - @teknium1)
See every model's reasoning, then watch the answer stream in - When a MoA ensemble runs, each reference model's full output now renders as its own labelled block - you can read what GPT-5 thought, what Claude thought, and what Grok thought, before the aggregator synthesizes them into one answer. And that final answer now streams to you live instead of appearing all at once after a long silence. This works in the CLI, the TUI, and the desktop app. You get to watch the committee deliberate, not just read the verdict. (#53793, #53855, #55625, #56101 - @teknium1)
The agent verifies its own work - "done" means proven, not claimed - Hermes now records verification evidence for coding work and can decide it's finished by actually running your project's checks, not by asserting success. /goal gained completion contracts: you state what "done" looks like, and the standing-goal loop judges completion against that evidence instead of stopping when the model feels like it. There's a pre_verify hook for wiring in custom checks and a one-time migration that tunes the defaults sensibly. The difference between "I think I fixed it" and "the tests pass, here's proof." (#50501, #52285, #55413, #53552 - @teknium1, @OutThisLife)
/learn - turn anything into a reusable skill by describing it - Run /learn and Hermes distills a reusable skill out of whatever you point it at - a directory, a URL, or just the workflow you walked it through five minutes ago. It writes the skill to the standards in your CONTRIBUTING.md automatically. The next time you need that workflow, it's already there. Teaching Hermes a new trick is now a single command, not a manual skill-authoring session. (#51506, #52372 - @teknium1)
/journey - a playable timeline of everything Hermes has learned about you - The CLI and TUI gained /journey, a learning timeline that shows the memories and skills Hermes has accumulated over time - and you can edit or delete any of them right from the view. Pair it with the desktop's new memory graph (a top-down, playable radial timeline of memories and skills) and for the first time you can actually see what your agent knows, watch it grow, and prune what's wrong. Your agent's memory stops being a black box. (#55555, #55859, #55226 - @OutThisLife)
Delegate a pile of work and keep going - background fan-out - delegate_task can now fan out multiple subagents that all run in the background: your chat is never blocked, and when every subagent finishes, their results come back as a single consolidated turn. Kick off "research these five competitors in parallel" or "audit these three modules," then carry on with something else while a small fleet works. When it's all done, you get one clean summary instead of babysitting each one. (#49734 - @teknium1)
First-class coding Projects in the desktop app - The desktop app gained real, per-profile Projects - a sidebar of your codebases, a coding rail, a review pane, git worktree management, and agent-facing project tools, all backed by a proper project repo lane model. Instead of scattered chat sessions, your coding work is organized into projects the agent understands and can act on. It's the desktop turning into an actual coding cockpit. (#49037, #54385, #54517 - @OutThisLife)
Run Hermes at scale - scale-to-zero and drain coordination - The gateway can now go dormant when idle and quiesce cleanly before a restart, migration, or auto-update - without dropping in-flight conversations. A hosted or relay-only Hermes can scale to zero when nobody's talking to it and wake back up on demand, and disruptive lifecycle actions coordinate an external drain so nobody gets cut off mid-turn. Running Hermes for a team or as a hosted service just got a lot more production-grade. (#52243, #52937, #54824 - @teknium1, @benbarclay)
Cheaper self-improvement - smarter background review - The post-turn self-improvement fork (the one that decides whether to save a memory or skill) now routes to an auxiliary model, digests context instead of replaying the whole conversation, and adapts its cadence - so the "learn from what just happened" loop that runs after your turns costs a fraction of what it used to. You keep the self-improvement, you stop paying full main-model price for it. (#49252 - @teknium1)
Compose your next prompt in your editor - /prompt - /prompt opens your $EDITOR so you can hand-write a long, multi-line prompt in real markdown instead of fighting a one-line input box. Draft a detailed spec, a structured question, or a big paste, save, and it's queued as your next message. Small thing, huge quality-of-life win for anyone who writes Hermes more than a sentence at a time. (#50509 - @teknium1)
Google Vertex AI - Gemini through your GCP service account, no static key - Vertex AI is now a first-class provider for Gemini models over Vertex's OpenAI-compatible endpoint. The reason a plain custom-provider setup always died mid-session is that Vertex has no static API key - every request needs a short-lived OAuth2 access token (~1h TTL) minted from a service-account JSON or Application Default Credentials. Hermes now mints and auto-refreshes those tokens for you, so if your org runs Gemini through Google Cloud, you point Hermes at your service account and it just works - no token-pasting, no mid-session expiry. (#56363 - @teknium1, @slawt)
Security round - This window hardened several surfaces: MCP-config persistence attack surface locked down, cron base_url overrides that could exfiltrate provider credentials blocked, a non-reusable sentinel for prefix secrets in file reads, Slack app-level (xapp-) token redaction, a browser cloud-metadata floor enforced on every backend, and an aiohttp CVE floor across the lazy messaging paths. Fewer ways for a prompt-injected or misconfigured session to leak a credential. (#50476, #56196, #54166, #56227, #52349, #56237 - <a class="user-mention notranslate" data-hovercard-type="user" data-hovercard-url="/users/teknium1/hovercard" data-octo-click="hovercard-link-click" data-octo-dimensions="link_type:se
