Data

Tools

Indexes

Applications

Experiments

Awesome

An open API service indexing awesome lists of open source software.

https://github.com/ddpie/agent-studio

基于 AWS Bedrock AgentCore 的 AI Agent 编排平台，通过自然语言创建、管理和对话 | AI agent orchestration on AWS Bedrock AgentCore - create, manage, and chat with agents via natural language
https://github.com/ddpie/agent-studio

Last synced: 8 days ago
JSON representation

基于 AWS Bedrock AgentCore 的 AI Agent 编排平台，通过自然语言创建、管理和对话 | AI agent orchestration on AWS Bedrock AgentCore - create, manage, and chat with agents via natural language

Awesome Lists containing this project

README 

          
# Agent Studio



基于 AWS Bedrock AgentCore 的 Agent 编排平台。用自然语言向 Meta-Agent 描述需求，它为你创建、部署、维护可直接运行的 Agent。Meta-Agent 由 Kiro CLI 驱动，Agent 基于 Strands。



## 能做什么



- **对话创建 Agent** — 告诉 Meta-Agent 你的需求，它写 prompt、挑工具、生成代码、完成部署

- **可视化调试** — 对话流式返回，每次 tool 调用的输入输出都显示在消息流里

- **Agent 互调** — `link_agent` 把一个 Agent 挂为另一个的工具，通过 A2A 协议调用，密钥自动下发

- **知识库 (RAG)** — 对话式创建 KB，上传文档自动向量化（S3 Vectors + Cohere Multilingual v3），Agent 通过 `kb_retrieve` 按需检索

- **Skill 热插拔** — AgentSkills.io 格式，运行时按需加载，跨 Agent 复用

- **多模态输入** — 图片 / PDF / Excel / CSV / TSV 自动解析，无需额外配置

- **定时触发** — 可视化 cron 构建器，每次执行留痕为卡片，可直接查看运行详情

- **MCP 工具集** — AWS 官方 MCP 目录全量接入，按需启用；per-target IAM 角色，最小权限

- **Workspace IAM 隔离** — 三层权限模型（Permission Boundary 天花板 → per-workspace 角色 → target 级授权）；Admin Console 一键授权，敏感 target 二次确认；`SimulatePrincipalPolicy` 实时检查，进度条展示每个 target 的已授权/缺失 action

- **跨会话记忆** — 基于 AgentCore Memory，Agent 自动记住用户偏好与事实，跨 session、跨设备持续生效；用户可在记忆抽屉中查看和删除

- **平台级可观测性** — 调用追踪、延迟分位数、错误率、Token 成本按 Agent 和 workspace 自动聚合，一屏总览

- **Channel 集成** — Agent 接入 IM 平台，用户直接在聊天工具中对话，流式卡片回复；零公网端点暴露（全部出站 WebSocket）。已支持飞书，Slack 和钉钉计划中

- **Marketplace** — 跨 workspace 发布和克隆 Agent / Skill / Tool，元数据公开，源码克隆后才可见



## 应用场景示例



一个"什么都会"的 Agent 理想丰满，实践脆弱。工具多了选不准，逻辑一复杂牵一发动全身，出了问题难以定位具体环节。这和单体应用的困境如出一辙：拆分、专注、协作才是出路。业务团队通常面临两个选择：忍受单体 Agent 复杂度膨胀带来的脆弱，或者等工程团队逐个开发、联调、部署垂类 Agent。



| 场景 | 描述 | 链接 |

|------|------|------|

| 游戏内容工厂 | 用自然语言编排验证第三条路：运营团队通过对话搭建 9-Agent 流水线（创作→审核→翻译），挂载世界观知识库，自动拦截版本泄露和设定冲突，不写代码 | [docs/demos/game-content-factory](docs/demos/game-content-factory/) |



## Meta-Agent 能力总览



Meta-Agent 内置 47 个工具，覆盖 Agent 从创建到运维的完整生命周期：



| 能力域 | 具体能力 |

|--------|---------|

| **Agent 全生命周期** | 创建、更新、验证（`validate_agent`）、部署、归档、恢复、永久删除；支持 zip 和 harness 两种运行时 |

| **Skill 系统** | 创建、编辑脚本文件、从 URL/Markdown 导入、挂载到 Agent、跨 Agent 同步更新、删除 |

| **多 Agent 编排** | A2A 链接/取消链接、调用其他 Agent、创建定时调度 |

| **知识库 (RAG)** | 创建 KB、上传文档、检查摄取状态、挂载/卸载到 Agent、删除 |

| **MCP 工具接入** | 浏览 50+ AWS 官方 MCP Server、查看 target 工具列表、workspace 权限校验后一键接入 |

| **运维 & 调试** | 实时日志查看、调用 trace 分析、打包代码预览、Agent Secrets 管理、workspace 权限检查 |

| **Prompt 工程** | 自动生成符合规范的 system prompt（约束分层、反模式预防、工具-Prompt 联动检查） |

| **内置工具库** | 浏览预构建工具模板（web_search、s3_read、chart 等），优先复用再自定义 |



所有操作通过自然语言对话触发，无需记忆工具名或参数。



## 两种 Agent 运行时



| 运行时 | 适合 | 不支持 |

|---|---|---|

| **zip**(默认) | 需要自定义 Python 工具、MCP、Skills、A2A 关联、浏览器/代码解释器 | — |

| **harness**(实验性) | 只需要 prompt + 模型 + 长期记忆,看重**创建可靠性**和**最短冷启动** | MCP、自定义工具、Skills、linked agents |



两者共享聊天、Memory、可观测、成本页面等所有平台能力;区别只在 Agent 执行层。



## 架构



```mermaid

%%{init: {'flowchart': {'nodeSpacing': 20, 'rankSpacing': 30}} }%%

flowchart LR

    User((用户)) --> Web[Web Console]

    User --> Ch[Channel
飞书 ...]

    Web & Cron[定时器] & Ch --> WS



    subgraph WS["Workspace · Agent 编排"]

        direction LR

        Meta[Meta-Agent] -.编排.-> A1 & A2 & A3

        A1[Agent A] -->|A2A| A2[Agent B]

        A3[Agent C]

        A1 ~~~ A3

    end



    subgraph Res["Workspace 资源"]

        direction TB

        S[Skills]

        K[(知识库)]

        M[MCP]

        Se[Secrets]

    end



    subgraph AC[AgentCore]

        direction TB

        CI[Code Interpreter]

        B[Browser]

        Me[Memory]

        O[Observability]

    end



    LLM[LLMs · Bedrock]



    WS <--> Res

    WS --> AC

    WS --> LLM

```



完整系统图（CloudFront / Lambda / EventBridge / Evaluator 等）与关键设计说明见 [docs/architecture.md](docs/architecture.md)。



## Workspace 协作模型



**Workspace** 是 Agent Studio 的协作单元。一个 workspace 下的成员共享 Agent、Skill、知识库和 MCP 工具配置，协同开发和调用 Agent。每个 workspace 拥有独立的 IAM 角色和资源隔离边界，workspace 之间数据互不可见。通过 Marketplace 可跨 workspace 发布和克隆公开资产。



## 安全与权限



Agent Studio 对每个 workspace 实施三层权限控制，在赋能 Agent 访问 AWS 资源的同时防止越权：



```

Permission Boundary（天花板）

  └─ Workspace Role（身份策略）

       └─ Target Grants（按需授权）

```



- **Permission Boundary** — `AgentStudioWorkspaceCeiling` 管理策略封顶 workspace 角色的最大权限。写操作、IAM 变更、角色提权被显式 Deny，无论身份策略怎么配都不会超出天花板

- **Per-workspace 角色** — 按需创建 `AgentStudio-ws-{id}` IAM 角色并绑定 Permission Boundary。未创建角色的 workspace 只能使用平台内置工具（搜索、图表、代码解释器）

- **Target 级授权** — 管理员在 Admin Console 按 MCP target 逐项授权（如 CloudWatch、DynamoDB、IAM）。IAM 策略从 `mcp-registry.yaml` 单一数据源自动生成，消除手工维护的漂移风险

- **敏感标识与二次确认** — 每个 target 标注 sensitivity 等级（low / medium / high）。medium 以上的授权操作弹出确认弹窗，展示风险原因和具体 action 列表

- **实时权限检查** — `SimulatePrincipalPolicy` 实时模拟，每个 target 展示进度条和已授权/缺失 action 明细。Agent 创建时 `validate_agent` 自动校验 MCP 权限



## 部署



### 前置条件



- 已配置凭据的 AWS 账号：`aws sts get-caller-identity` 能正常返回

- 使用的 IAM 身份能够 `cdk bootstrap`，并能创建 IAM 角色、策略与 Permission Boundary

- 本机装有 **Node.js 18+**、**Python 3.10+**、**Docker**（CDK Lambda bundling 需要 Docker daemon 在跑）；以及 `npm`、`jq`、`yq`——脚本会做版本预检



### 一键部署



首次部署两步走。先准备 `.env`：



```bash

cp .env.example .env

```



打开 `.env` 填入三项必填：



- `AGENT_STUDIO_REGION` — `us-east-1` 或 `us-west-2`

- `AGENT_STUDIO_ACCOUNT_ID` — 12 位 AWS 账号 ID

- `ORIGIN_VERIFY_SECRET` — 用 `openssl rand -hex 32` 生成



其它字段（`META_AGENT_ID`、`COGNITO_*`、`CLOUDFRONT_*`）部署脚本会自动回填。



然后一键部署，约 15–20 分钟：



```bash

bash scripts/deploy-all.sh --region us-east-1   # 或 --region us-west-2

```



脚本依次完成：CDK 基础设施（Cognito / Lambda / CloudFront / WAF / DDB / S3）→ Base zip → Meta-Agent → 前端 build 与同步 → CloudFront invalidation。重复执行时只更新有变化的部分。



### 常用子命令



```bash

bash scripts/deploy-all.sh --only-agent       # 仅更新 Meta-Agent

bash scripts/deploy-all.sh --only-frontend    # 仅更新前端

bash scripts/deploy-all.sh --skip-frontend    # 基础设施 + Meta-Agent

bash scripts/deploy-all.sh --dry-run          # 预检 + cdk diff



bash scripts/build-mcp.sh                     # 构建 MCP Docker 镜像

bash scripts/deploy-mcp.sh                    # 部署 MCP targets（按 region 自动过滤不可用项）



cd frontend && npm run dev                    # 本地开发

```



### 常见问题



- **`AGENT_STUDIO_REGION is not set`** — Vite 不再静默 fallback，必须在 `.env` 里显式写明

- **Docker not running** — CDK 打包 Lambda 需要 Docker daemon；`--only-frontend` / `--only-agent` / `--dry-run` 不依赖 Docker



## 项目结构



```

agent-studio/

├── frontend/              # React 19 + Vite 8 + Tailwind 4 + Zustand 5

├── meta-agent/            # Meta-Agent on AgentCore Runtime（Kiro CLI 驱动）

│   ├── kiro_adapter/      # Kiro 胶水层（ACP 驱动、stdio MCP、SSE mapper）

│   ├── tools/             # Meta 工具集（Agent 生命周期、skill、MCP、schedule、secret、link 等）

│   ├── tools_library/     # 预构建 Agent 工具模板（web_search、s3_read、chart 等）

│   └── templates/         # 代码生成 + prompt 模板

├── lambda/

│   ├── crud/              # Python CRUD Lambda（含 kiro_key.py 管理 key + usage）

│   ├── invoke-node/       # Node.js SSE streaming proxy

│   ├── a2a-proxy/         # A2A JSON-RPC proxy

│   └── channels/          # IM Channel relay（飞书/Slack/钉钉 WebSocket）+ worker

├── mcp-runtime/           # AWS MCP 目录接入（mcp-registry.yaml 控制启用项）+ Dockerfile

├── infra/                 # AWS CDK (TypeScript)

├── scripts/               # 部署 / 测试脚本

└── .env.example

```



## 测试



```bash

bash scripts/run-tests.sh

```



---



# Agent Studio (English)



An agent orchestration platform on AWS Bedrock AgentCore. Describe what you need in natural language and the Meta-Agent creates, deploys, and maintains the agent for you. The Meta-Agent is powered by the Kiro CLI; agents run on Strands.



## What It Does



- **Conversational agent creation** — Tell the Meta-Agent what you need; it writes the prompt, picks tools, generates the code, deploys.

- **Visualized debugging** — Streaming responses with every tool call's input and output surfaced in the chat stream.

- **Agent-to-agent** — `link_agent` mounts one agent as another's tool over A2A; credentials are provisioned automatically.

- **Knowledge Base (RAG)** — Create KBs via chat, upload documents for automatic vectorization (S3 Vectors + Cohere Multilingual v3), agents retrieve via `kb_retrieve` on demand.

- **Hot-swappable skills** — AgentSkills.io format, lazy-loaded at runtime, reusable across agents.

- **Multimodal input** — Images / PDF / Excel / CSV / TSV parsed out of the box, no extra setup.

- **Scheduled triggers** — Visual cron builder; every run is archived as a card with a click-through to full run details.

- **MCP toolbelt** — The full AWS-official MCP catalog, enable what you need; per-target IAM roles, least privilege.

- **Workspace IAM isolation** — Three-layer permission model (Permission Boundary ceiling → per-workspace role → target-level grants); Admin Console one-click grants with sensitivity badges and confirmation dialogs; `SimulatePrincipalPolicy` real-time checks with progress bars per target.

- **Cross-session memory** — Powered by AgentCore Memory: agents remember user preferences and facts across sessions and devices. Users can view and manage memories from the chat drawer.

- **Platform-level observability** — Trace timeline, latency percentiles, error rates, and token costs aggregated per agent and per workspace on a single screen.

- **Channel integration** — Connect agents to IM platforms; users chat directly with streaming card replies. Zero public endpoints (all outbound WebSocket). Feishu shipped; Slack and DingTalk planned

- **Marketplace** — Publish and clone agents / skills / tools across workspaces; metadata is public, source stays private until cloned.



## Example Scenarios



A "do-everything" agent sounds ideal but proves fragile in practice. Too many tools lead to selection errors, complex logic creates cascading failures, and when something breaks it's hard to pinpoint which step went wrong. This mirrors the monolith problem in software: split, focus, collaborate is the way out. Teams typically face two choices: tolerate the brittleness of a bloated monolith agent, or wait for engineering to build, integrate, and deploy specialized agents one by one.



| Scenario | Description | Link |

|------|------|------|

| Game Content Factory | A third way validated through natural-language orchestration: ops teams build a 9-agent pipeline (create→review→translate) via conversation, backed by a lore knowledge base that auto-blocks version leaks and lore conflicts — no code required | [docs/demos/game-content-factory](docs/demos/game-content-factory/) |



## Meta-Agent Capabilities



The Meta-Agent ships with 47 built-in tools covering the full agent lifecycle:



| Domain | Capabilities |

|--------|-------------|

| **Agent lifecycle** | Create, update, validate (`validate_agent`), deploy, archive, restore, permanently delete; supports both zip and harness runtimes |

| **Skill system** | Create, edit script files, import from URL/Markdown, attach to agents, sync updates across agents, delete |

| **Multi-agent orchestration** | A2A link/unlink, invoke other agents, create scheduled triggers |

| **Knowledge Base (RAG)** | Create KB, upload documents, check ingestion status, attach/detach from agents, delete |

| **MCP integration** | Browse 50+ AWS-official MCP servers, inspect target tool lists, workspace permission check before enabling |

| **Ops & debugging** | Live logs, invocation trace analysis, assembled code preview, agent secrets management, workspace permission audit |

| **Prompt engineering** | Auto-generate standards-compliant system prompts (constraint layering, anti-pattern prevention, tool–prompt sync checks) |

| **Built-in tool library** | Browse pre-built tool templates (web_search, s3_read, chart, …); prefer reuse over custom code |



All operations are triggered through natural-language conversation — no need to memorize tool names or parameters.



## Architecture



```mermaid

%%{init: {'flowchart': {'nodeSpacing': 20, 'rankSpacing': 30}} }%%

flowchart LR

    User((User)) --> Web[Web Console]

    User --> Ch[Channel
Feishu ...]

    Web & Cron[Scheduled trigger] & Ch --> WS



    subgraph WS["Workspace · Agent Orchestration"]

        direction LR

        Meta[Meta-Agent] -.orchestrate.-> A1 & A2 & A3

        A1[Agent A] -->|A2A| A2[Agent B]

        A3[Agent C]

        A1 ~~~ A3

    end



    subgraph Res["Workspace Resources"]

        direction TB

        S[Skills]

        K[(Knowledge Base)]

        M[MCP]

        Se[Secrets]

    end



    subgraph AC[AgentCore]

        direction TB

        CI[Code Interpreter]

        B[Browser]

        Me[Memory]

        O[Observability]

    end



    LLM[LLMs · Bedrock]



    WS <--> Res

    WS --> AC

    WS --> LLM

```



Full system diagram (CloudFront / Lambda / EventBridge / Evaluator, …) and key design notes live in [docs/architecture.md](docs/architecture.md).



## Workspace Collaboration



A **Workspace** is the unit of collaboration in Agent Studio. Members of a workspace share agents, skills, knowledge bases, and MCP tool configurations, collaborating on agent development and invocation. Each workspace has its own IAM role and resource isolation boundary — data is invisible across workspaces. The Marketplace enables cross-workspace publishing and cloning of public assets.



## Security & Permissions



Agent Studio enforces a three-layer permission model on every workspace, enabling AWS resource access while preventing privilege escalation:



```

Permission Boundary (ceiling)

  └─ Workspace Role (identity policy)

       └─ Target Grants (on-demand)

```



- **Permission Boundary** — The `AgentStudioWorkspaceCeiling` managed policy caps the maximum privileges of any workspace role. Write operations, IAM mutations, and privilege escalation are explicitly denied

- **Per-workspace role** — `AgentStudio-ws-{id}` IAM roles are created on demand and bound to the Permission Boundary. Workspaces without a role can only use platform built-in tools (search, charts, code interpreter)

- **Target-level grants** — Admins grant permissions per MCP target (e.g. CloudWatch, DynamoDB, IAM) from the Admin Console. IAM policies are auto-generated from `mcp-registry.yaml` as the single source of truth, eliminating manual drift

- **Sensitivity labels & confirmation** — Each target is labeled low / medium / high sensitivity. Medium+ grants trigger a confirmation dialog showing risk reasons and the specific IAM actions being granted

- **Real-time checks** — `SimulatePrincipalPolicy` provides live simulation with progress bars and per-action granted/missing status. `validate_agent` automatically verifies MCP permissions at agent creation time



## Deployment



### Prerequisites



- An AWS account with working credentials — `aws sts get-caller-identity` must succeed

- An IAM identity allowed to run `cdk bootstrap` and to create IAM roles, policies, and a Permission Boundary

- Local tooling: **Node.js 18+**, **Python 3.10+**, **Docker** (CDK Lambda bundling requires a running daemon), plus `npm`, `jq`, `yq` — the script preflight-checks all of them



### One-Click Deploy



First time? Two steps. Start with `.env`:



```bash

cp .env.example .env

```



Open `.env` and fill in three required fields:



- `AGENT_STUDIO_REGION` — `us-east-1` or `us-west-2`

- `AGENT_STUDIO_ACCOUNT_ID` — your 12-digit AWS account ID

- `ORIGIN_VERIFY_SECRET` — generate with `openssl rand -hex 32`



The rest (`META_AGENT_ID`, `COGNITO_*`, `CLOUDFRONT_*`) is filled in automatically by the deploy script.



Then deploy — 15–20 minutes end-to-end:



```bash

bash scripts/deploy-all.sh --region us-east-1   # or --region us-west-2

```



The script runs: CDK infra (Cognito / Lambda / CloudFront / WAF / DDB / S3) → base zip → Meta-Agent → frontend build & sync → CloudFront invalidation. Re-runs only update what actually changed.



### Common Subcommands



```bash

bash scripts/deploy-all.sh --only-agent       # Meta-Agent code only

bash scripts/deploy-all.sh --only-frontend    # Frontend only

bash scripts/deploy-all.sh --skip-frontend    # Infra + Meta-Agent

bash scripts/deploy-all.sh --dry-run          # Preflight + cdk diff



bash scripts/build-mcp.sh                     # Build MCP Docker images

bash scripts/deploy-mcp.sh                    # Deploy MCP targets (unavailable ones auto-filtered)



cd frontend && npm run dev                    # Local dev

```



### Troubleshooting



- **`AGENT_STUDIO_REGION is not set`** — Vite no longer silently falls back; set it explicitly in `.env`.

- **Docker not running** — CDK Lambda bundling needs a running Docker daemon. `--only-frontend` / `--only-agent` / `--dry-run` do not.



## Project Structure



```

agent-studio/

├── frontend/              # React 19 + Vite 8 + Tailwind 4 + Zustand 5

├── meta-agent/            # Meta-Agent on AgentCore Runtime (Kiro-driven)

│   ├── kiro_adapter/      # Kiro glue (ACP driver, stdio MCP, SSE mapper)

│   ├── tools/             # Meta tools (agent lifecycle, skills, MCP, schedule, secrets, link, ...)

│   ├── tools_library/     # Pre-built agent tool templates (web_search, s3_read, chart, ...)

│   └── templates/         # Codegen + prompt templates

├── lambda/

│   ├── crud/              # Python CRUD Lambda (includes kiro_key.py — key + usage)

│   ├── invoke-node/       # Node.js SSE streaming proxy

│   ├── a2a-proxy/         # A2A JSON-RPC proxy

│   └── channels/          # IM channel relay (Feishu/Slack/DingTalk WebSocket) + worker

├── mcp-runtime/           # AWS MCP catalog integration (mcp-registry.yaml gates what's enabled) + Dockerfile

├── infra/                 # AWS CDK (TypeScript)

├── scripts/               # Deploy + test scripts

└── .env.example

```



## Testing



```bash

bash scripts/run-tests.sh

```