Hermes Agent, the open-source AI runtime from Nous Research, has released v0.13.0 (v2026.5.7) — “The Tenacity Release” — its most ambitious update to date. The release ships just one week after v0.12.0 “Curator” — and it arrives in the midst of a rapidly maturing open-source agent ecosystem and represents 864 commits, 588 merged PRs, 829 files changed, 128,366 insertions, and 282 issues closed (13 of them P0 severity). The project has also crossed 138,390 GitHub stars, up from 135K just two days ago.
“The Tenacity Release — Hermes Agent now finishes what it starts.” — Teknium1, release notes
🧩 Multi-Agent Kanban: Delegate to an AI Team That Actually Finishes
Hermes Agent sits among the top open-source AI agent frameworks, and this release cements that position.
The headline feature of v0.13.0 is the durable multi-agent Kanban board system — a complete reimplementation that lets you spin up a board, drop tasks on it, and have multiple Hermes worker agents pick them up, hand off, and close them out with real reliability guarantees.
Worker Lifecycle That Doesn’t Forget
The Kanban system solves the hardest problem in multi-agent orchestration: making sure tasks actually complete. Key durability mechanisms include:
- Heartbeat tracking — workers pulse their status; silent workers get reclaimed
- Zombie detection — processes that crash without cleanup are detected and their tasks re-assigned (#21183)
- Per-task max_retries — override retry budgets at the individual task level (#21330)
- Hallucination gate + recovery — detects when a worker fabricates card completions and triggers recovery UX (#20232)
- Auto-block on incomplete exit — workers that exit without completing their task automatically self-block (#21214)
- Diagnostics engine — a generic distress-signal engine for task health monitoring (#20332)
One install, many kanbans. Your AI team now holds itself accountable.
The dashboard also gained per-platform home-channel notification toggles (#19864), multi-project board support (#19653), and workspace-aware inline creation forms (#19679).
🎯 /goal — The Agent Doesn’t Forget What You Asked
The new /goal slash command introduces the “Ralph loop” as a first-class primitive: lock the agent onto a target and it stays on task across turns, even when you switch context or the session bounces (#18262). This is a fundamental shift from stateless turn-by-turn interaction to persistent objective-oriented execution.
💾 Checkpoints v2 — State Persistence Rewritten
Checkpoints received a ground-up rewrite with real pruning, disk guardrails, and no more orphan shadow repos (#20709). Combined with auto-resume after gateway restart (#21192), sessions now survive /update restarts, source-file reloads, and gateway bounces seamlessly.
🛡️ Security Wave — 8 P0 Closures
v0.13.0 delivers the most aggressive security hardening pass in Hermes Agent’s history. (For context on the broader state of AI agent development as of May 2026, see our comprehensive overview.) Eight Priority-0 issues were closed:
| Security Fix | Impact |
|---|---|
| Redaction ON by default | Secrets are now auto-redacted in all output (#21193) |
| Discord role-allowlists guild-scoped | Closes CVSS 8.1 cross-guild DM bypass (#21241) |
| WhatsApp rejects strangers by default | No more responding to unknown numbers (#21291) |
| MCP OAuth TOCTOU fix | Time-of-check/time-of-use window closed (#21176) |
| auth.json TOCTOU fix | Credential writers now atomic (#21194) |
| Browser SSRF floor | Cloud-metadata endpoints blocked in hybrid routing (#21228) |
hermes debug share redaction |
Logs redacted before upload (#19318) |
| Cron prompt-injection scanning | Skill content scanned before execution (#21350) |
Additional security improvements include SRI integrity for dashboard plugin scripts, 0600 permissions on all credential files, and platform allowlists (allowed_channels / allowed_chats / allowed_rooms) across Slack, Telegram, Mattermost, Matrix, and DingTalk (#21251).
🌐 Language, Platforms & Providers
Seven i18n Locales
The static gateway and CLI messages now translate to Chinese, Japanese, German, Spanish, French, Ukrainian, and Turkish. The docs site also gained a zh-Hans (Chinese) locale (#20231, #20329, #20467, #20474).
Google Chat — Platform #20
Google Chat joins the supported platform roster — the 20th messaging platform. It ships alongside a generic platform-plugin hooks surface so third-party adapters can drop in without touching core code (#21306, #21331).
Pluggable Providers
The ProviderProfile ABC and plugins/model-providers/ directory make inference providers a pluggable surface — third-party providers can now be dropped in without modifying the runtime (#20324).
🎬 New Capabilities
video_analyzetool — native video understanding on Gemini and compatible multimodal models (#19301)- xAI Custom Voices — voice cloning as a TTS provider (#18776)
- Post-write delta lint — the agent now lints its own
write_fileandpatchcalls for Python, JSON, YAML, and TOML syntax errors (#20191) no_agentcron mode — script-only watchdog pattern for cron jobs (#19709)transform_llm_outputplugin hook — reshape or filter LLM output before it hits the conversation (#21235)- ACP
/steerand/queue— direct in-flight agents or queue follow-ups from Zed, VS Code, or JetBrains (#18114) - API server
X-Hermes-Session-Key— long-term memory scoping per session (#20199)
🔌 New Models & Skills
New models include deepseek/deepseek-v4-pro, x-ai/grok-4.3, openrouter/owl-alpha (free tier), tencent/hy3-preview, and Arcee Trinity Large with temperature + compression overrides.
6 new optional skills landed: Shopify (Admin + Storefront GraphQL), here.now location awareness, shop-app personal shopping assistant, an Anthropic financial-services bundle, kanban-video-orchestrator, and searxng-search.
📊 By the Numbers
| Metric | Value |
|---|---|
| GitHub Stars | 138,390 |
| Forks | 21,332 |
| Open Issues | 8,904 |
| Commits since v0.12.0 | 864 |
| Merged PRs | 588 |
| Community contributors | 295 |
| P0 issues closed | 13 |
The Takeaway
v0.13.0 “Tenacity” marks a decisive inflection point for Hermes Agent. The multi-agent Kanban system transforms it from a single-agent CLI tool into a durable multi-agent orchestration platform. The /goal persistence, Checkpoints v2, and gateway auto-resume together create a system that genuinely remembers and finishes what it starts. And the 8-P0 security wave — combined with platform allowlists, i18n, and pluggable providers — signals that Nous Research is investing heavily in making Hermes Agent enterprise-ready.
With 138K stars and nearly 300 community contributors in a single weekly cycle, the project’s trajectory shows no signs of slowing down. If you haven’t tried Hermes Agent since the early days, v0.13.0 is the release to come back for.
Full changelog: v2026.4.30…v2026.5.7