# Hive Framework — Outcome-Driven Multi-Agent Harness

> Hive Framework turns natural-language goals into generated agent graphs and dashboards, giving teams a production-minded multi-agent harness.

## Install

Save as a script file and run:

## Quick Use

1. Clone and run the guided setup:
   ```bash
   git clone https://github.com/aden-hive/hive.git
   cd hive
   ./quickstart.sh
   ```
2. On Windows use:
   ```powershell
   .\quickstart.ps1
   ```
3. Verify:
   - Open the browser dashboard, create one sample agent, and run it once end-to-end.

## Intro

Hive Framework turns natural-language goals into generated agent graphs and dashboards, giving teams a production-minded multi-agent harness.

- **Best for:** teams that want graph generation, dashboards, and MCP-connected business workflows in one harness
- **Works with:** Python 3.11+, uv workspaces, browser dashboard, LiteLLM-compatible providers, MCP tool integrations
- **Setup time:** 20-40 minutes

## Practical Notes

- Quant: the repo advertises 102 MCP tools and a uv workspace split across core runtime and tool environments.
- Quant: the quickstart provisions at least two virtual environments (`core/.venv` and `tools/.venv`) plus encrypted credentials.

## Why it matters

Hive is notable because it is opinionated about the harness layer: goal intake, graph generation, execution, observation, and evolution are treated as one loop.

- The quickstart is explicit that `pip install -e .` is the wrong path, which reduces a common failure mode in workspace-style Python repos.
- The dashboard-first UX is useful when non-framework specialists need to inspect how agents were generated and run.
- The project emphasizes business-system connectivity through MCP, which is the main difference between a demo agent and a production workflow.

## Rollout pattern

- Use the quickstart script exactly as documented and avoid improvising the install path on first run.
- Pilot one outcome-driven workflow that touches a single external system before you let Hive generate broader graphs.
- Treat the browser dashboard as an audit surface: capture screenshots or logs from the first few generated runs.

## Watchouts

Because Hive generates orchestration structures for you, the main risk is over-trusting generated graphs before you understand the tool permissions and business-system side effects.

### FAQ

**Q: Can I install it with pip editable mode?**
A: No. The README warns that the project uses a uv workspace layout and that `pip install -e .` will not work correctly.

**Q: What makes it different from a basic agent SDK?**
A: It includes graph generation, dashboarding, credential handling, and MCP-oriented system integration.

**Q: What is the first validation step?**
A: Run the quickstart, open the dashboard, and execute one sample agent with limited scope.

## Source & Thanks

> Source: https://github.com/aden-hive/hive
> License: Apache-2.0
> GitHub stars: 10,302 · forks: 5,647

---

<!-- ZH -->

## 快速使用

1. 克隆并运行向导式安装：
   ```bash
   git clone https://github.com/aden-hive/hive.git
   cd hive
   ./quickstart.sh
   ```
2. Windows 请使用：
   ```powershell
   .\quickstart.ps1
   ```
3. 验证：
   - 打开浏览器仪表盘，创建一个示例 Agent，并完整执行一次。

## 简介

Hive Framework 试图把自然语言目标自动变成 Agent 图与控制台，为团队提供更偏生产环境的多 Agent harness，减少继续手写、维护和排查零散编排脚本所带来的长期负担。

- **适合谁：** 希望把图式编排、仪表盘与 MCP 业务连接统一在一个 harness 中的团队
- **可搭配：** Python 3.11+、uv workspace、浏览器仪表盘、LiteLLM 兼容模型与 MCP 工具集成
- **准备时间：** 20-40 分钟

## 实战建议

- 量化信息：仓库首页直接标出 102 个 MCP tools，并把运行时与工具环境拆成 uv workspace。
- 量化信息：quickstart 至少会准备 `core/.venv`、`tools/.venv` 和加密凭证存储。 

## 为什么值得收录

Hive 的独特点在于它明确把“harness 层”当成产品主体：目标输入、图生成、执行、观测和演化是同一个闭环。

- README 明确提醒不要用 `pip install -e .`，这能避免很多 uv workspace 项目的常见安装坑。
- 以浏览器仪表盘为中心的交互方式，更适合让非框架开发者也能看懂 Agent 是怎么生成和运行的。
- 项目强调通过 MCP 连接业务系统，这正是 Demo Agent 与生产工作流之间的关键差别。

## 落地路径

- 第一次部署按 quickstart 原样执行，不要自作主张改安装路径。
- 先试 1 个只连接单一外部系统的 outcome-driven workflow，再扩展更复杂图。
- 把浏览器仪表盘当成审计界面，前几次运行务必保留截图或日志证据。 

## 注意事项

因为它会替你生成编排结构，最大的风险是你还没搞清工具权限和业务副作用，就过度信任自动生成的 Agent 图。

### FAQ

**可以直接用 pip editable 安装吗？**
答：不行。README 已明确说明它是 uv workspace，`pip install -e .` 会导致功能异常。

**它和普通 Agent SDK 的区别是什么？**
答：它把图生成、仪表盘、凭证管理和 MCP 系统连接一起打包了。

**第一步怎么验证？**
答：跑 quickstart，打开 dashboard，再执行一个权限有限的示例 Agent。

## 来源与感谢

> Source: https://github.com/aden-hive/hive
> License: Apache-2.0
> GitHub stars: 10,302 · forks: 5,647


---
Source: https://tokrepo.com/en/workflows/hive-framework-outcome-driven-multi-agent-harness
Author: Agent Toolkit