ADK-Rust：43 个 crate 拆开来用，社区复刻的 Rust Agent 运行时

在挑 Agent 运行时的时候翻到了这个。

ADK-Rust 是 zavora-ai 社区维护的 Rust Agent 开发框架，不是 Google 官方项目，但名字和 API 风格跟 Google ADK 对齐（GitHub topic 里标了 google-adk-rust）。当前版本 v2.2.0，要求 Rust 1.95+，Apache-2.0 协议，43 个可独立发布的 crate 按职责拆开，用哪块引哪块。截至发稿 667 stars，crates.io 全工作空间合计超过 50 万次下载。

仓库：github.com/zavora-ai/adk-rust

---

为什么值得看

ADK 生态里不缺 Python 实现，缺的是敢拿 Rust 真正把 Agent 运行时生产化的项目。ADK-Rust 在几处正好压在关键点上：

• 模块化到 crate 级：43 个 crate，每个都能独立发布、独立版本。写 CLI 脚本只引 adk-core + adk-agent，上 HTTP 服务再加 adk-server，不用把整个框架带进来。
• Agent loop 开销 568 μs：跟 Python SDK（253 μs）在一个数量级，比 LangGraph（1,228 ms）快了两千倍。但 Python SDK 的循环开销反而更低——Rust 的收益主要体现在冷启动（109 ms vs 501 ms）和内存（~15 MB vs 92.7 MB）。
• adk-skill crate：能解析 SKILL.md 格式的 Agent Skills，做词法匹配和提示注入——对用 Claude Code / Codex 等工具链的人直接有用。
• graph 工作流的 durable resume：SQLite checkpointer，进程崩了再启动可以从断点恢复，不是假的持久化。

---

架构：43 个 crate 分四层

README 把功能按 tier 分成四档，加进 Cargo.toml 的时候直接选：

[dependencies]
adk-rust = "2.2.0"
# adk-rust = { version = "2.2.0", features = ["standard"] }   # +server/auth/graph/eval
# adk-rust = { version = "2.2.0", features = ["enterprise"] }  # +realtime/browser/RAG
# adk-rust = { version = "2.2.0", features = ["full"] }        # 全部

Tier	包含
minimal（默认）	Gemini、agent、runner、sessions
standard	minimal + OpenAI/Anthropic、tools、memory、telemetry、server、auth、graph、eval
enterprise	standard + realtime、browser、RAG、payments、AWP
full	enterprise + audio、代码执行、sandbox

Tier 是起点不是上限——features = ["minimal", "audio"] 可以在 minimal 上单独加音频能力，不必整体升级到 enterprise。

---

最小可运行示例

use adk_rust::prelude::*;
use adk_rust::Launcher;

#[tokio::main]
async fn main() -> AnyhowResult<()> {
    dotenvy::dotenv().ok();
    let model = GeminiModel::new(&std::env::var("GOOGLE_API_KEY")?, "gemini-3.7-flash")?;

    let agent = LlmAgentBuilder::new("assistant")
        .instruction("You are a helpful assistant. Be concise and accurate.")
        .model(Arc::new(model))
        .build()?;

    Launcher::new(Arc::new(agent)).run().await?;
    Ok(())
}

换 provider 只换 client，agent 和 tools 不动：

Provider	客户端构造	Feature Flag
Gemini	GeminiModel::new(key, "gemini-3.7-flash")	默认
OpenAI	OpenAIClient::new(OpenAIConfig::new(key, model))	openai
Anthropic	AnthropicClient::new(AnthropicConfig::new(key, model))	anthropic
DeepSeek	DeepSeekClient::chat(key)	deepseek
Ollama	OllamaModel::new(OllamaConfig::new(model))	ollama
Bedrock	BedrockClient::new(...).await?	bedrock

还支持 xAI Grok、Mistral、Fireworks、Together AI 等 OpenAI 兼容预设，以及 mistral.rs 本地推理（Gemma 4、Qwen 3.5）。

---

graph 工作流：durable resume 怎么用

这是我觉得最值钱的部分。adk-graph 实现了 LangGraph 风格的有向图调度，叠了几个关键能力：

Checkpointing（持久化）：

• 内存 checkpointer（测试用）
• SQLite checkpointer（生产可用）
• Delta checkpointer（只存变化量，节省空间）

Durable resume：进程重启后从数据库恢复，只需要共享同一个 SQLite 文件：

let checkpointer = SqliteCheckpointer::new("agent_state.db").await?;
let graph = MyGraph::builder()
    .checkpointer(checkpointer)
    .build()?;

// 崩了重启，同一个 thread_id 继续
let run = graph.resume_or_start(thread_id, input).await?;

Human-in-the-loop：图节点可以 pause 等待人工确认，审批内容绑到 digest，审批的和实际执行的是同一份内容：

graph.add_node("sensitive_action", sensitive_node)
    .require_approval(ApprovalPolicy::DigestBound)

with_goto 动态路由：节点在运行时决定自己的下一个节点，不需要预先声明边，适合 LLM 输出决定下一步的场景。

Time travel：可以回到历史 checkpoint 重新执行，用于调试或对比不同分支。

---

adk-skill：解析 SKILL.md

adk-skill crate 专门做 AgentSkills 解析，这对现在用 Claude Code / Codex 工具链的人直接有用：

use adk_skill::SkillIndex;

let index = SkillIndex::discover("/path/to/.claude/skills").await?;
let matches = index.match_input("generate a banner for this article");
// → 返回 banner-creator skill 的 SKILL.md 内容和触发置信度

词法匹配（不需要嵌入模型），找到匹配后自动注入 prompt。支持 .skills 目录发现和索引，能扫整个 ~/.claude/skills/ 树。

---

关键 crate 索引

挑几个有工程价值的：

Crate	干什么
adk-graph	LangGraph 风格图调度，SQLite checkpoint，durable resume，time travel
adk-skill	SKILL.md 解析 + 词法匹配 + prompt 注入，支持 .skills 目录发现
adk-realtime	OpenAI Realtime + Gemini Live，双向音频/视频，VAD，情感对话
adk-computer-use	受管桌面自动化，digest 绑定审批中断，篡改无效
adk-sandbox	进程/WASM 沙箱，macOS Seatbelt，Linux bubblewrap
adk-memory	语义检索 + bi-temporal 知识图谱
adk-rag	文档切片 + 嵌入 + 向量检索 + reranking，6 种后端
adk-audio	STT/TTS，Deepgram 流式，ONNX 本地（Whisper/Moonshine/Kokoro）
adk-payments	ACP/AP2 适配器，可审计支付流，durable journal
adk-devtools	read_file/write_file/bash 等 DevToolset，沙箱隔离 workspace

---

脚手架工具

cargo install cargo-adk

cargo adk new my-agent                       # 基础 Gemini agent
cargo adk new my-agent --template graph      # graph 工作流 + checkpoint
cargo adk new my-agent --template realtime   # 实时语音 agent
cargo adk new my-agent --template api        # HTTP 服务
cargo adk new my-agent --template agent-engine  # Gemini Enterprise BYOC
cargo adk new my-agent --addon mcp --addon guardrails  # 叠 addon

生成的项目带嵌入式 UI，打开 http://127.0.0.1:8080/ui/ 可以看到对话流、工具结果、workflow 拓扑图、事件 timeline 和 OpenTelemetry tracing。

---

性能数据

用 cargo adk bench 在 Apple M 系列 + macOS + gemini-2.5-flash 上测，同一负载：

框架	冷启动	Agent Loop 均值	P95	峰值 RSS
ADK-Rust	109 ms	568 μs	615 μs	~15 MB
Gemini Python SDK	501 ms	253 μs	334 μs	69.7 MB
LangGraph	502 ms	1,228 ms	1,228 ms	92.7 MB

值得注意的是：Agent loop 均值 Python SDK（253 μs）比 ADK-Rust（568 μs）低——Rust 的主要优势在冷启动（快 4.6x）和内存（低 4.6x），在 loop 开销上跟 Python SDK 是同数量级但并非更快。LangGraph 的 1,228 ms 是另一个量级，跟这两个不在同一个对比维度上。

自报数据，请打折看。用 cargo adk bench --dry-run 可以先估成本再跑。

---

v2.2.0 新增的 Gemini Enterprise 路径

v2.2.0 完成了 Gemini Enterprise Agent Platform 的消费路径，全部可选、可组合：

• Gen AI Evaluation Service bridge
• Vertex AI RAG Engine 检索与接地
• Agent Retrieval 向量存储
• Agent Registry 发现与注册
• Skill Registry 远程 skill 加载
• 远程 ReasoningEngine agent 可作为 sub-agent 调用

Graph 工作流新增原生工具确认暂停。Tracing 修复了一次调用导出多条断裂 trace 的问题。

---

几点局限

• 社区维护，非 Google 官方：API 和 Google ADK 对齐，但不是官方实现，稳定性保障不同。
• NOASSERTION license：GitHub API 返回的是 NOASSERTION，README 标注 Apache-2.0，商用前需自行核查 LICENSE 文件。
• macOS sandbox 更完善：Windows AppContainer 沙箱未实现，在文档里明确写了。
• Python SDK loop 开销更低：如果你的 agent 是 loop-heavy 而非 process-heavy，Python SDK 反而更快。
• adk-managed 和 adk-codeact-monty 标为 Experimental：生产慎用。

---

配套仓库

• adk-ui：动态 UI 生成（github.com/zavora-ai/adk-ui）
• adk-studio：可视化 agent builder（github.com/zavora-ai/adk-studio）
• adk-playground：120+ 可运行示例（github.com/zavora-ai/adk-playground）

Podcast 系列（Episode 1–3）是用 ADK-Rust 自己的音频能力生成的——adk-audio crate 驱动 Chirp3-HD 多说话人 TTS，脚本 + 幻灯片 + 音频片段拼成视频，零人工录音。

开源代码仅供学习研究，用于生产前请自行评估稳定性和 license。

---

仓库：github.com/zavora-ai/adk-rust
版本：v2.2.0 | Stars：667 | License：Apache-2.0 | Rust：1.95+

When shopping for an agent runtime, I came across this one.

ADK-Rust is a Rust agent development framework maintained by the zavora-ai community organization — not a Google official project, but the name and API style align with Google ADK (the GitHub topics include google-adk-rust). Current version is v2.2.0, requires Rust 1.95+, Apache-2.0 license. 43 independently publishable crates split by responsibility — pull in only what you need. 667 stars at time of writing, 500K+ cumulative crates.io downloads across the workspace.

Repository: github.com/zavora-ai/adk-rust

---

Why it’s worth looking at

The ADK ecosystem has no shortage of Python implementations. What’s missing is a project that seriously productionizes an agent runtime in Rust. ADK-Rust hits several points that matter:

• Modular to the crate level: 43 crates, each independently publishable and versioned. A CLI script pulls in only adk-core + adk-agent. An HTTP service adds adk-server. No need to drag the full framework in.
• 568 μs agent loop overhead: Same order of magnitude as Python SDK (253 μs), two thousand times faster than LangGraph (1,228 ms). That said, Python SDK’s loop overhead is actually lower — Rust’s advantage is cold start (109 ms vs 501 ms) and memory (~15 MB vs 92.7 MB).
• adk-skill crate: Parses SKILL.md-format Agent Skills, does lexical matching and prompt injection — directly useful for anyone using Claude Code or Codex toolchains.
• Graph workflow with durable resume: SQLite checkpointer. Process crashes and restarts resume from the breakpoint. Not simulated persistence.

---

Architecture: 43 crates in four tiers

Features are organized into tiers, selected directly in Cargo.toml:

[dependencies]
adk-rust = "2.2.0"
# adk-rust = { version = "2.2.0", features = ["standard"] }   # +server/auth/graph/eval
# adk-rust = { version = "2.2.0", features = ["enterprise"] }  # +realtime/browser/RAG
# adk-rust = { version = "2.2.0", features = ["full"] }        # everything

Tier	Includes
minimal (default)	Gemini, agent, runner, sessions
standard	minimal + OpenAI/Anthropic, tools, memory, telemetry, server, auth, graph, eval
enterprise	standard + realtime, browser, RAG, payments, AWP
full	enterprise + audio, code execution, sandbox

A tier is a starting point, not a ceiling. features = ["minimal", "audio"] adds audio on top of minimal without upgrading to enterprise.

---

Minimal runnable example

use adk_rust::prelude::*;
use adk_rust::Launcher;

#[tokio::main]
async fn main() -> AnyhowResult<()> {
    dotenvy::dotenv().ok();
    let model = GeminiModel::new(&std::env::var("GOOGLE_API_KEY")?, "gemini-3.7-flash")?;

    let agent = LlmAgentBuilder::new("assistant")
        .instruction("You are a helpful assistant. Be concise and accurate.")
        .model(Arc::new(model))
        .build()?;

    Launcher::new(Arc::new(agent)).run().await?;
    Ok(())
}

Swap the provider by swapping the client — the agent and tools are unchanged:

Provider	Client	Feature
Gemini	GeminiModel::new(key, "gemini-3.7-flash")	default
OpenAI	OpenAIClient::new(OpenAIConfig::new(key, model))	openai
Anthropic	AnthropicClient::new(AnthropicConfig::new(key, model))	anthropic
DeepSeek	DeepSeekClient::chat(key)	deepseek
Ollama	OllamaModel::new(OllamaConfig::new(model))	ollama
Bedrock	BedrockClient::new(...).await?	bedrock

Also supports xAI Grok, Mistral, Fireworks, Together AI, and other OpenAI-compatible presets, plus mistral.rs for local inference (Gemma 4, Qwen 3.5).

---

Graph workflows: how durable resume works

This is the most valuable part. adk-graph implements LangGraph-style directed graph scheduling with several critical capabilities:

Checkpointing:

• In-memory checkpointer (for tests)
• SQLite checkpointer (production-ready)
• Delta checkpointer (stores only deltas, saves space)

Durable resume: recover from a database after process restart, sharing only an SQLite file:

let checkpointer = SqliteCheckpointer::new("agent_state.db").await?;
let graph = MyGraph::builder()
    .checkpointer(checkpointer)
    .build()?;

// Crashed and restarted — same thread_id continues
let run = graph.resume_or_start(thread_id, input).await?;

Human-in-the-loop: graph nodes can pause waiting for human approval, bound to a digest — what you approved is what runs:

graph.add_node("sensitive_action", sensitive_node)
    .require_approval(ApprovalPolicy::DigestBound)

with_goto dynamic routing: a node decides its own successor at runtime without pre-declared edges — good for LLM-driven control flow.

Time travel: rewind to a historical checkpoint and re-execute, for debugging or branch comparison.

---

adk-skill: parsing SKILL.md

The adk-skill crate specifically handles AgentSkills parsing — directly useful for Claude Code and Codex toolchain users:

use adk_skill::SkillIndex;

let index = SkillIndex::discover("/path/to/.claude/skills").await?;
let matches = index.match_input("generate a banner for this article");
// → returns banner-creator skill's SKILL.md content and trigger confidence

Lexical matching (no embedding model needed). Finds matches and auto-injects prompts. Supports .skills directory discovery and indexing, can scan an entire ~/.claude/skills/ tree.

---

Key crates

A selection of the ones with engineering value:

Crate	Purpose
adk-graph	LangGraph-style graph scheduling, SQLite checkpoint, durable resume, time travel
adk-skill	SKILL.md parsing + lexical matching + prompt injection, .skills directory discovery
adk-realtime	OpenAI Realtime + Gemini Live, bidirectional audio/video, VAD, affective dialogue
adk-computer-use	Governed desktop automation, digest-bound approval interrupts, tamper-evident
adk-sandbox	Process/WASM sandbox, macOS Seatbelt, Linux bubblewrap
adk-memory	Semantic retrieval + bi-temporal knowledge graph
adk-rag	Chunking + embeddings + vector search + reranking, 6 backends
adk-audio	STT/TTS, Deepgram streaming, ONNX local models (Whisper/Moonshine/Kokoro)
adk-payments	ACP/AP2 adapters, auditable payment flows, durable journals
adk-devtools	read_file/write_file/bash DevToolset, sandboxed workspace

---

Performance numbers

Measured with cargo adk bench on Apple M-series + macOS + gemini-2.5-flash (self-reported — apply a discount):

Framework	Cold Start	Loop Overhead (mean)	P95	Peak RSS
ADK-Rust	109 ms	568 μs	615 μs	~15 MB
Gemini Python SDK	501 ms	253 μs	334 μs	69.7 MB
LangGraph	502 ms	1,228 ms	1,228 ms	92.7 MB

Note: Python SDK’s loop overhead (253 μs) is lower than ADK-Rust (568 μs). Rust’s main advantage is cold start (4.6x faster) and memory (4.6x less). LangGraph’s 1,228 ms is a different order of magnitude entirely.

Run cargo adk bench --dry-run to see cost estimates before running.

---

Key limitations

• Community-maintained, not Google official: API aligns with Google ADK, but stability guarantees differ from an official implementation.
• NOASSERTION license: GitHub’s API returns NOASSERTION; README badges say Apache-2.0. Verify the LICENSE file before commercial use.
• Windows AppContainer sandbox not implemented: Explicitly documented. macOS Seatbelt and Linux bubblewrap work; Windows doesn’t.
• Python SDK has lower loop overhead: For loop-heavy (not process-heavy) agents, Python SDK is actually faster.
• adk-managed and adk-codeact-monty are Experimental: Avoid in production.

---

Related repos

• adk-ui: Dynamic UI generation (github.com/zavora-ai/adk-ui)
• adk-studio: Visual agent builder (github.com/zavora-ai/adk-studio)
• adk-playground: 120+ runnable examples (github.com/zavora-ai/adk-playground)

The podcast series (Episodes 1–3) is generated by ADK-Rust itself — adk-audio drives Chirp3-HD multi-speaker TTS, and the script + slide deck + audio segments are concatenated with ffmpeg into a video. Zero manual voice recording.

Open-source code is for learning and research. Evaluate stability and license before production use.

---

Repository: github.com/zavora-ai/adk-rust
Version: v2.2.0 | Stars: 667 | License: Apache-2.0 | Rust: 1.95+