Just 11 days after the Foundation Release, Nous Research has shipped Hermes Agent v0.15.0 — “The Velocity Release” — and followed it with a same-day v0.15.1 hotfix on May 29. If Foundation was about breadth (install anywhere, connect to anything), Velocity is about depth — making everything dramatically faster, smaller, and smarter.
The stats are staggering:
“The Velocity Release — Hermes gets dramatically faster — to start, to run, to ship work, and to grow.” — Teknium1, release notes
Here is the full breakdown of what lands in v0.15.0 and v0.15.1.
🔧 The Big Refactor: run_agent.py Goes from 16K to 3.8K Lines
The headline engineering story of v0.15.0 is the surgical dismantling of the monolithic run_agent.py — the file at the heart of Hermes’ agent conversation loop. It has been reduced from 16,083 lines to 3,821 lines (-76%), with the extracted logic redistributed across 14 cohesive modules under agent/.
Every extraction keeps a thin forwarder for backward compatibility, so behavior is unchanged — but the codebase is now dramatically easier to navigate, test, and contribute to. This is the kind of infrastructure debt payment that pays compounding interest for every future release.
A 16,000-line file is a research prototype. A 3,800-line module tree is a production platform.
🧩 Kanban: From Feature to Multi-Agent Platform
The Kanban system — first introduced in v0.13.0 “Tenacity” — has matured from a task board into a full multi-agent orchestration platform across 104 merged pull requests. New capabilities include:
- Orchestrator auto-decomposition — the Kanban master breaks complex goals into sub-tasks automatically
- Swarm topology — workers self-organize into dependency-aware execution graphs
- Scheduled tasks — cron-integrated board slots that fire on a timer
- Worktree-per-task — each task gets an isolated working directory to prevent cross-contamination
- Per-task model overrides — assign different LLMs to different board slots
This positions Kanban as a genuine rival to dedicated multi-agent frameworks like AutoGen and CrewAI — but running inside a single hermes install with zero extra infrastructure.
⚡ Cold-Start Performance: Another Second Shaved Off
The cold-start performance wave that began in Foundation continues. v0.15.0 delivers:
- ~1 more second off
hermeslaunch — boot-to-prompt is now dramatically faster - 47% fewer per-conversation function calls — the agent spends less time on boilerplate, more time doing work
hermes --versionbenchmark win — Hermes Agent now beats Codex CLI head-to-head in cold-start benchmarks
For an agent runtime that users invoke dozens of times per day, every second compounds into real productivity.
🔍 session_search: 4,500× Faster and Free
The session_search skill — which lets Hermes search across all your past conversations — has been rebuilt. The result: 4,500× faster lookups at zero additional cost. No vector database, no external API, no embedding service required. This is pure algorithmic optimization on the local conversation store.
🛡️ Promptware Defense: Protection Against Brainworm-Class Attacks
v0.15.0 ships promptware defense — a security layer specifically designed to detect and neutralize Brainworm-class prompt injection attacks. In an era where agents reading untrusted web content is table stakes, this is a critical addition. The defense layer:
- Scans incoming content for known injection patterns
- Sandboxes suspicious content before it reaches the agent loop
- Reports attempted attacks without disrupting legitimate work
🔐 Bitwarden Secrets Manager: One Token to Rule Them All
One of the most practical quality-of-life improvements: Bitwarden Secrets Manager integration. Instead of managing N different API keys for N different providers (Claude, Grok, OpenAI, etc.), you now configure a single Bitwarden bootstrap token, and Hermes retrieves the rest automatically at runtime. This is a massive win for:
- Security — fewer tokens in plaintext config files
- Portability — move your Hermes install without migrating keys
- Team deployments — share secrets via Bitwarden without sharing files
🎨 Ink TUI: Multi-Session Orchestrator
The terminal user interface gets its own multi-session orchestrator, letting you manage multiple Hermes sessions from a single TUI dashboard. Switch between conversations, inspect Kanban boards, and monitor background tasks — all without leaving the terminal.
📦 Skill Bundles: One Slash Command, One Workflow
A new skill bundle system lets you group multiple skills behind a single slash command. Instead of loading skills one at a time, you define a bundle — @deploy could load k8s-skills, docker-skills, monitoring-skills, and rollback-skills in one shot.
🌐 Platform Expansion
v0.15.0 extends Hermes’ already impressive platform reach:
- ntfy becomes the 23rd messaging platform — push notifications to any device via ntfy.sh
- Two new
image_genproviders: Krea 2 (Medium + Large) and a FAL plugin port - Nous-approved MCP catalog with an interactive picker — discover and install MCP servers from within Hermes
- OpenHands orchestration skill — run OpenHands workflows as Hermes skills
- Deep xAI integration round — Web Search plugin via Grok,
xai-oauthproxy upstream, retired-May-15 model detection withhermes migrate xai, natural TTS speech-tag pauses,base_urlleak guard, and OpenAI-style execution guidance for Grok
🪲 v0.15.1 Hotfix (May 29) — The Dashboard Infinite-Reload Fix
Released just hours after v0.15.0, v0.15.1 is a same-day hotfix release with 28 commits, 21 merged PRs, and 9 contributors. The headline fix:
- Dashboard 401 reload loop fixed — In loopback mode (Docker, hosted Hermes, fresh installs), the dashboard’s identity probe (
/api/auth/me) returns 401 by design, but v0.15.0’s stale-token reload guard treated every 401 as a rotated session and full-page-reloaded. This caused an infinite loop — captured in issues #34206 and #34202, fixed in PR #30698 by @austinpickett.
Additional fixes in v0.15.1:
- Docker
--insecureas explicit env opt-in — no longer derived from bind host - Kanban worker SIGTERM handling improved
/modelpicker unification across UI and CLI/yolosession bypass for power users- Full 19,932-entry skills.sh catalog now loads correctly
.mdmedia delivery restoration in conversations- Gateway probe-stepdown safety — graceful degradation on upstream failure
- Web-URL redaction passthrough fixed for private browsing mode
- Kanban worker vision on referenced images — workers can now see attached images
- Hindsight observation set as default behavioral mode
- MCP server bare-command PATH resolution on systems with custom PATHs
- arm64 PR-build cache fixes for Apple Silicon and ARM Linux contributors
📊 The Growth Trajectory
172,224 GitHub Stars · 28,932 Forks · 14,855 Open Issues
Since the Foundation release (May 16): +17K stars (+11%) in 13 days. The project is now the #1 most-starred AI agent runtime on GitHub by a wide margin, and the velocity of contributions shows no signs of slowing.
The Bottom Line
Hermes Agent v0.15.0 “Velocity” is not a feature release — it’s a foundational investment in speed and maintainability. The run_agent.py refactor alone will pay dividends for every release that follows. The Kanban platform maturation, session_search rewrite, promptware defense layer, and Bitwarden integration together create a runtime that’s faster, safer, and more practical for daily production use.
The same-day v0.15.1 hotfix demonstrates the project’s operational maturity — shipping a critical bug fix within hours of a major release, with clear communication and detailed changelogs. That’s the hallmark of a project that treats its users seriously.
With 172K stars, 28.9K forks, and a 321-contributor release cycle, Hermes Agent is not just the fastest-growing open-source agent runtime — it’s becoming the default choice for developers who want an agent that ships real work, at real speed.
Stay tuned for our next deep dive — the Velocity release has already generated over 200 new open issues, and the community sprint shows no signs of stopping.