Hermes Agent v0.13.0: The Tenacity Release

Hermes Agent v0.13.0 的关键词是 Tenacity。放到产品语境里，它想解决的不是“再加一个炫技工具”，而是 Agent 在真实使用里最容易掉链子的地方：任务变长、入口变多、中途会断、状态要接得上。

先看结论：这次最值得关注的 8 个变化

DurabilityKanban 多 Agent 看板

把任务变成可追踪、可恢复、可交接的持久队列。

Goals/goal 持久目标

让 Hermes 跨轮次继续推进一个目标，不用你反复说“继续”。

GatewayGateway 自动恢复

网关重启、更新或中断后，尽量恢复被打断的会话和路由。

Cronno_agent 模式

定时脚本可以不调用 LLM，零 token 做监控告警。

Videovideo_analyze

把视频分析纳入 Agent 工具链，适合产品 demo、教程、会议片段。

VoicexAI Custom Voices

TTS 支持 xAI 自定义声音和 voice cloning。

PlatformGoogle Chat

Google Chat 成为 Hermes 的第 20 个 messaging platform。

Extensibility安全和插件化增强

密钥脱敏默认开启，模型 provider 可以用插件方式扩展。

1. Kanban：把多 Agent 协作变成一块可持续的任务板

Kanban 是这次最重要的能力之一。以前你让 Agent 拆任务，很多时候是在一次对话里临时派生一个子任务。它更像函数调用：发出去，等结果，失败了就失败了。

v0.13.0 的 Kanban 更像一个持久工作队列。任务、状态、评论、依赖、worker 日志都会落在本地数据库和工作区里。一个 worker 崩了，可以 reclaim；任务没做完退出，会自动 block；多次失败可以触发 retry 上限；长任务可以 heartbeat。

初始化并启动 dispatcher

hermes kanban init
hermes gateway start

创建任务

hermes kanban create "研究 Hermes v0.13.0 的所有更新" \
  --assignee researcher \
  --workspace scratch \
  --max-retries 3

创建多项目看板

hermes kanban boards create launch-research \
  --name "Launch Research" \
  --description "Hermes v0.13.0 article workflow" \
  --switch

查看任务状态

hermes kanban list
hermes kanban show <task_id>
hermes kanban watch
hermes kanban runs <task_id>
hermes kanban log <task_id>

如果你在聊天界面或消息平台里，也可以直接用 slash command：

聊天里驱动 Kanban

/kanban list
/kanban create "写一篇 v0.13.0 更新文章" --assignee writer
/kanban comment t_abcd "重点写使用方法，不要只复述 release note"
/kanban dispatch --max 3

不要把 Kanban 理解成 Jira 替代品。

它更准确的定位是：Hermes 内部的多 Agent 工作队列。它解决的是 Agent 任务如何被拆分、恢复、交接和审计，不是完整团队项目管理系统。

2. `/goal`：再也不用一直说“继续”

/goal 让你给当前会话设定一个 standing goal。每轮结束后，Hermes 会用一个轻量 judge 判断目标是否完成；如果没完成，就自动把继续工作的话题送回同一个会话。

设定一个持久目标

/goal 修复当前项目所有失败测试，并确认 pytest 全部通过

命令	作用
`/goal <text>`	设定或替换当前 standing goal，并立即开始执行。
`/goal status`	查看当前目标、状态和已使用轮次。
`/goal pause`	暂停自动继续，但保留目标。
`/goal resume`	恢复自动继续，并重置轮次计数。
`/goal clear`	清除目标。

配置 goal 轮次预算

goals:
  max_turns: 20

为 goal judge 配辅助模型

auxiliary:
  goal_judge:
    provider: openrouter
    model: google/gemini-3-flash-preview

/goal 适合不会一轮完成的事，比如修完整个测试集、迁移一组文件、连续整理资料、补完一篇长文。它不是“永久记忆”，也不是完全自主规划系统；它的重点是让一个明确目标可以跨轮次保持。

3. Cron `no_agent`：定时任务不一定要叫醒 LLM

很多自动化任务其实不需要 Agent 思考。比如内存过高就提醒、磁盘快满就报警、CI 失败就发一段日志。这类任务最适合 v0.13.0 的 script-only cron，也就是 --no-agent。

创建一个脚本

mkdir -p ~/.hermes/scripts

cat > ~/.hermes/scripts/disk-alert.sh <<'EOF'
#!/usr/bin/env bash
df -h / | awk 'NR > 1 && $5+0 >= 90 { print "Disk " $5 " full on " $6 }'
EOF

chmod +x ~/.hermes/scripts/disk-alert.sh

创建 no-agent cron

hermes cron create "*/15 * * * *" \
  --no-agent \
  --script disk-alert.sh \
  --deliver telegram \
  --name "disk-alert"

脚本有 stdout，就把 stdout 发出去；stdout 为空，就静默。全程不调用模型。它的效果不是“更聪明”，而是更轻、更稳定、更便宜。

4. Gateway auto-resume 与 Checkpoints v2：中断后更容易接上

v0.13.0 强化了会话和状态持久性。Gateway 在中途重启、/update 重启、源文件 reload 等情况下，会尝试在回来后恢复被打断的会话。

Checkpoints v2 则重写了状态持久化，强调 pruning、磁盘 guardrails，并减少旧版本中 orphan shadow repo 这类维护问题。

查看 checkpoint 状态

hermes checkpoints status
hermes checkpoints prune

5. 视频与语音：让 Agent 看得懂更多输入，也能输出声音

`video_analyze`

v0.13.0 新增了 video_analyze 工具，用于分析视频 URL 或本地视频文件。它需要视频能力模型，比如 Gemini 或其他兼容的多模态模型。

启用 video toolset

hermes -t video

视频分析提示词示例

请分析 /absolute/path/demo.mp4。
按时间线列出视频里展示的产品功能，并指出哪些片段适合剪成发布素材。

xAI Custom Voices

TTS 这次支持 xAI Custom Voices，可以把 xAI Console 里创建的自定义 voice 用作 Hermes 的语音输出。

配置 xAI TTS

tts:
  provider: xai
  xai:
    voice_id: "your_custom_voice_id"
    language: "en"
    sample_rate: 24000

6. Google Chat 与 SearXNG：入口和搜索后端都更可配置

Google Chat

Google Chat 是这次新增的消息平台。它使用 Cloud Pub/Sub pull subscription 接收入站事件，用 Chat REST API 发送回复。对 Google Workspace 用户来说，这意味着 Hermes 可以更自然地进入团队协作入口。

安装 Google Chat extra

pip install 'hermes-agent[google_chat]'

基础环境变量

GOOGLE_CHAT_PROJECT_ID=my-chat-bot-123
GOOGLE_CHAT_SUBSCRIPTION_NAME=projects/my-chat-bot-123/subscriptions/hermes-chat-events-sub
GOOGLE_CHAT_SERVICE_ACCOUNT_JSON=/home/you/.hermes/google-chat-sa.json
GOOGLE_CHAT_ALLOWED_USERS=you@yourdomain.com

SearXNG

SearXNG 可以作为免费、自托管、无 API key 的搜索后端，但它只负责搜索，不负责网页正文抽取。推荐搭配方式是：SearXNG 负责搜索，Firecrawl、Tavily、Exa 或 Parallel 负责抽取。

SearXNG 搜索加 Firecrawl 抽取

web:
  search_backend: "searxng"
  extract_backend: "firecrawl"

7. 安全、Dashboard 与插件化：更像一个可管理的本地系统

v0.13.0 的 security wave 关闭了多项高优先级问题。对长期运行 Gateway、接入聊天平台、启用 cron 和外部工具的人来说，这些更新尤其重要。

Secret redaction 默认开启。
hermes debug share 上传前会做脱敏。
WhatsApp 默认拒绝陌生人。
Discord role allowlist 变成 guild-scoped。
Slack、Telegram、Mattermost、Matrix、DingTalk 等平台增强 allowlist。
Cron 会扫描 assembled prompt，包括 skill content，降低 prompt injection 风险。

启动 Dashboard

pip install 'hermes-agent[web,pty]'
hermes dashboard

模型 provider 插件目录

~/.hermes/plugins/model-providers/<name>/

安全改进不等于绝对安全。

更准确的说法是：默认值更保守，更多高风险入口有了防护和限制。只要你把 Agent 接到消息平台、浏览器、MCP、cron 或文件系统，仍然应该认真维护 allowlist、权限边界和日志保留策略。

最短上手路径

1. 升级和检查

hermes update -y
hermes --version
hermes doctor

2. 打开中文静态提示

hermes config set display.language zh

3. 试一个 /goal

/goal 整理当前项目 README，补充安装、更新和常见问题，并确认链接可用

4. 试 Kanban

hermes kanban init
hermes gateway start
hermes kanban create "写一份项目改进建议" --assignee researcher --triage
hermes kanban specify --all

如果你只是想快速了解这次更新，优先看 Kanban、/goal、Google Chat 和 Cron no_agent。它们最能代表 v0.13.0 的方向：让 Hermes Agent 从一次对话工具，继续往可持续工作的 Agent 系统靠近。

参考来源

Hermes Agent v0.13.0 is called The Tenacity Release for a reason. The main story is not a single flashy tool. It is about the places where agents usually struggle in real use: long tasks, multiple entry points, interrupted runs, and state that needs to survive.

The short version: 8 changes that matter most

DurabilityMulti-agent Kanban

A durable board for tasks that need tracking, recovery, handoff, and retry history.

GoalsPersistent /goal

Hermes can keep working toward a standing objective across turns without repeated “continue” prompts.

GatewayAuto-resume after restart

Interrupted sessions and routing state are more likely to come back after gateway restarts or updates.

Cronno_agent mode

Scheduled scripts can skip the LLM entirely and deliver stdout directly.

Videovideo_analyze

Video files and URLs can enter the agent workflow through video-capable multimodal models.

VoicexAI Custom Voices

TTS can use custom xAI voices, including voice cloning support.

PlatformGoogle Chat

Google Chat becomes Hermes Agent’s 20th messaging platform.

ExtensibilitySecurity and provider plugins

Secret redaction is on by default, and model providers can be added as plugins.

1. Kanban turns multi-agent work into a durable board

Kanban is one of the most important additions in this release. Traditional delegation behaves more like a function call: spawn a subtask, wait for the result, and lose most of the trail if things fail.

The new Kanban system is closer to a durable work queue. Tasks, status, comments, dependencies, worker logs, retry history, and handoff metadata live outside a single conversation context. A crashed worker can be reclaimed, incomplete exits can be blocked, and long tasks can heartbeat.

Initialize and start dispatcher

hermes kanban init
hermes gateway start

Create a task

hermes kanban create "Research every Hermes v0.13.0 update" \
  --assignee researcher \
  --workspace scratch \
  --max-retries 3

Create a project board

hermes kanban boards create launch-research \
  --name "Launch Research" \
  --description "Hermes v0.13.0 article workflow" \
  --switch

Inspect task state

hermes kanban list
hermes kanban show <task_id>
hermes kanban watch
hermes kanban runs <task_id>
hermes kanban log <task_id>

The same surface works from chat and messaging platforms through slash commands:

Drive Kanban from chat

/kanban list
/kanban create "Write a v0.13.0 update article" --assignee writer
/kanban comment t_abcd "Focus on usage, not just release-note summary"
/kanban dispatch --max 3

Do not read Kanban as a Jira replacement.

Its better framing is an internal Hermes work queue for multi-agent tasks: decomposition, recovery, handoff, and audit trail.

2. `/goal` removes the need to keep saying “continue”

/goal lets you set a standing objective for the current session. After each turn, Hermes uses a lightweight judge to decide whether the goal is complete. If not, it feeds a continuation prompt back into the same session.

Set a persistent goal

/goal Fix every failing test in the current project and verify pytest passes

Command	What it does
`/goal <text>`	Set or replace the standing goal and start immediately.
`/goal status`	Show the current goal, status, and turns used.
`/goal pause`	Pause auto-continuation while keeping the goal.
`/goal resume`	Resume the loop and reset the turn counter.
`/goal clear`	Clear the goal.

Configure the turn budget

goals:
  max_turns: 20

Route the judge to a cheaper auxiliary model

auxiliary:
  goal_judge:
    provider: openrouter
    model: google/gemini-3-flash-preview

Use it for tasks that should not stop after one turn: fixing a full test suite, migrating a set of files, iterating on research, or finishing a long-form deliverable. It is not permanent memory or a fully autonomous planning system; it is a way to keep a clear objective alive across turns.

3. Cron `no_agent`: scheduled jobs do not always need an LLM

Some automation does not require reasoning. Memory alerts, disk alerts, CI pings, and API polling often only need a script. That is exactly what script-only cron mode is for.

Create a script

mkdir -p ~/.hermes/scripts

cat > ~/.hermes/scripts/disk-alert.sh <<'EOF'
#!/usr/bin/env bash
df -h / | awk 'NR > 1 && $5+0 >= 90 { print "Disk " $5 " full on " $6 }'
EOF

chmod +x ~/.hermes/scripts/disk-alert.sh

Create a no-agent cron job

hermes cron create "*/15 * * * *" \
  --no-agent \
  --script disk-alert.sh \
  --deliver telegram \
  --name "disk-alert"

If the script prints stdout, Hermes delivers it. If stdout is empty, the run stays silent. No model call is made.

4. Gateway auto-resume and Checkpoints v2 make interruptions easier to recover from

v0.13.0 strengthens session and state durability. When the gateway restarts during a run, after /update, or after source reloads, Hermes is better at resuming the interrupted conversation and routing state.

Checkpoints v2 rewrites state persistence with real pruning and disk guardrails, reducing maintenance issues such as orphan shadow repositories.

Inspect checkpoint state

hermes checkpoints status
hermes checkpoints prune

5. Video and voice: richer inputs, richer outputs

`video_analyze`

The new video_analyze tool analyzes video URLs or local files through video-capable multimodal models such as Gemini or compatible providers.

Enable the video toolset

hermes -t video

Example video prompt

Analyze /absolute/path/demo.mp4.
List the product features shown in timeline order and identify clips suitable for launch material.

xAI Custom Voices

TTS now supports xAI Custom Voices, so a custom voice created in the xAI Console can be used for Hermes speech output.

Configure xAI TTS

tts:
  provider: xai
  xai:
    voice_id: "your_custom_voice_id"
    language: "en"
    sample_rate: 24000

6. Google Chat and SearXNG expand where Hermes can work

Google Chat

Google Chat is the new messaging platform in this release. It uses Cloud Pub/Sub pull subscriptions for inbound events and the Chat REST API for replies, making it useful for Google Workspace environments.

Install the Google Chat extra

pip install 'hermes-agent[google_chat]'

Basic environment variables

GOOGLE_CHAT_PROJECT_ID=my-chat-bot-123
GOOGLE_CHAT_SUBSCRIPTION_NAME=projects/my-chat-bot-123/subscriptions/hermes-chat-events-sub
GOOGLE_CHAT_SERVICE_ACCOUNT_JSON=/home/you/.hermes/google-chat-sa.json
GOOGLE_CHAT_ALLOWED_USERS=you@yourdomain.com

SearXNG

SearXNG can act as a free, self-hosted, no-key search backend. It is search-only, so pair it with a separate extraction provider such as Firecrawl, Tavily, Exa, or Parallel.

Use SearXNG for search and Firecrawl for extract

web:
  search_backend: "searxng"
  extract_backend: "firecrawl"

7. Security, Dashboard, and plugins make Hermes more manageable

The security wave in v0.13.0 closes multiple high-priority issues. These matter most if you run Gateway long-term, connect messaging platforms, enable cron, or expose external tools.

Secret redaction is on by default.
hermes debug share redacts before upload.
WhatsApp rejects strangers by default.
Discord role allowlists are scoped by guild.
Slack, Telegram, Mattermost, Matrix, and DingTalk gain stronger allowlist controls.
Cron scans assembled prompts, including skill content, for prompt injection risk.

Start Dashboard

pip install 'hermes-agent[web,pty]'
hermes dashboard

Model provider plugin directory

~/.hermes/plugins/model-providers/<name>/

Security improvements do not mean absolute safety.

The right reading is that defaults are more conservative and more high-risk surfaces have guardrails. You should still maintain allowlists, permission boundaries, and log retention policies.

Fast start path

1. Update and check

hermes update -y
hermes --version
hermes doctor

2. Enable Chinese static UI messages

hermes config set display.language zh

3. Try /goal

/goal Clean up the current project README, add install/update/FAQ sections, and verify links work

4. Try Kanban

hermes kanban init
hermes gateway start
hermes kanban create "Write a project improvement brief" --assignee researcher --triage
hermes kanban specify --all

If you only try a few things, start with Kanban, /goal, Google Chat, and Cron no_agent. Together they best represent the direction of v0.13.0: moving Hermes Agent from a single conversation tool toward a system that can sustain work.

这次更新，重点是让 Agent 更耐跑

This release is about making agents keep going

先看结论：这次最值得关注的 8 个变化

1. Kanban：把多 Agent 协作变成一块可持续的任务板

2. `/goal`：再也不用一直说“继续”

3. Cron `no_agent`：定时任务不一定要叫醒 LLM

4. Gateway auto-resume 与 Checkpoints v2：中断后更容易接上

5. 视频与语音：让 Agent 看得懂更多输入，也能输出声音

`video_analyze`

xAI Custom Voices

6. Google Chat 与 SearXNG：入口和搜索后端都更可配置

Google Chat

SearXNG

7. 安全、Dashboard 与插件化：更像一个可管理的本地系统

最短上手路径

参考来源

The short version: 8 changes that matter most

1. Kanban turns multi-agent work into a durable board

2. `/goal` removes the need to keep saying “continue”

3. Cron `no_agent`: scheduled jobs do not always need an LLM

4. Gateway auto-resume and Checkpoints v2 make interruptions easier to recover from

5. Video and voice: richer inputs, richer outputs

`video_analyze`

xAI Custom Voices

6. Google Chat and SearXNG expand where Hermes can work

Google Chat

SearXNG

7. Security, Dashboard, and plugins make Hermes more manageable

Fast start path

Sources

这次更新，重点是让 Agent 更耐跑

This release is about making agents keep going

先看结论：这次最值得关注的 8 个变化

1. Kanban：把多 Agent 协作变成一块可持续的任务板

2. /goal：再也不用一直说“继续”

3. Cron no_agent：定时任务不一定要叫醒 LLM

4. Gateway auto-resume 与 Checkpoints v2：中断后更容易接上

5. 视频与语音：让 Agent 看得懂更多输入，也能输出声音

video_analyze

xAI Custom Voices

6. Google Chat 与 SearXNG：入口和搜索后端都更可配置

Google Chat

SearXNG

7. 安全、Dashboard 与插件化：更像一个可管理的本地系统

最短上手路径

参考来源

The short version: 8 changes that matter most

1. Kanban turns multi-agent work into a durable board

2. /goal removes the need to keep saying “continue”

3. Cron no_agent: scheduled jobs do not always need an LLM

4. Gateway auto-resume and Checkpoints v2 make interruptions easier to recover from

5. Video and voice: richer inputs, richer outputs

video_analyze

xAI Custom Voices

6. Google Chat and SearXNG expand where Hermes can work

Google Chat

SearXNG

7. Security, Dashboard, and plugins make Hermes more manageable

Fast start path

Sources

2. `/goal`：再也不用一直说“继续”

3. Cron `no_agent`：定时任务不一定要叫醒 LLM

`video_analyze`

2. `/goal` removes the need to keep saying “continue”

3. Cron `no_agent`: scheduled jobs do not always need an LLM

`video_analyze`