Agently 4 🚀

更快构建生产级 AI 应用，实现稳定输出与可维护的工作流。

English Introduction | 中文介绍

---

🔥 最新文档 | 🚀 5分钟快速入门 | 💡 核心功能

---

官方 Skills

官方可安装的 Agently Skills 现已发布：

• GitHub: https://github.com/AgentEra/Agently-Skills
• 安装: npx skills add AgentEra/Agently-Skills

安装后您可以获得：

• 针对真实 Agently 工作的更好实现路径，从单次请求设计到 TriggerFlow、多智能体、MCP、工具、会话及 FastAPI 集成。
• 针对常见生产任务的开箱即用技能指南，无需在每次编码会话中重复解释框架概念。
• LangChain 和 LangGraph 的迁移指南，以便更快地将现有智能体系统转换为 Agently。
• 更高质量的编码智能体体验，因为 skills 包为智能体提供了具体的 Agently 模式、边界和推荐的实现路径。

---

📚 快速链接

• 文档 (EN): https://agently.tech/docs
• 文档 (中文): https://agently.cn/docs
• 智能体系统手册 (EN): https://agently.tech/docs/en/agent-systems/overview.html
• 智能体系统手册 (中文): https://agently.cn/docs/agent-systems/overview.html
• 官方 Agently Skills: https://github.com/AgentEra/Agently-Skills
• 安装 Agently Skills: npx skills add AgentEra/Agently-Skills

可选配套包：Agently Devtools

agently-devtools 是一个独立的配套包，用于运行时观察和开发者工具。

• 安装: pip install agently-devtools
• 依赖方向: agently-devtools -> agently
• 兼容性: agently-devtools 0.1.x 适用于 agently >=4.0.9

---

💡 核心功能概览

1) 📦 结构化数据输出

无需复杂的 prompt 工程即可获得稳定的 JSON 输出。

result = (
    agent.input("Introduce Python")
    .output({
        "name": (str, "语言名称"),
        "features": [(str, "特性")],
    })
    .start()
)

2) 🧩 TriggerFlow (事件驱动工作流)

使用简单的链式调用管理复杂的逻辑。

flow = TriggerFlow()

@flow.chunk
def step1(data):
    return "result_from_step1"

(
    flow.to(step1)
    .to(lambda d: print(d.value))
    .start("start_node")
)

3) 🔀 多模型切换 (无需重写代码)

通过简单的配置更改即可切换底座模型。

agent.set_settings("model", "gpt-4o")
# 切换模型只需修改一行配置
agent.set_settings("model", "deepseek-chat")

4) 💾 高级会话与上下文管理

灵活控制上下文窗口。

# 示例：仅保留最后 6 条上下文记录 (resize)
session = agent.activated_session
assert session is not None

def analysis_handler(full_context, context_window, memo, session_settings):
    if len(context_window) > 6:
        return "keep_last_six"
    return None

def keep_last_six(full_context, context_window, memo, session_settings):
    return None, list(context_window[-6:]), memo

session.register_analysis_handler(analysis_handler)
session.register_resize_handler("keep_last_six", keep_last_six)

5) 🔧 工具调用 + 日志

工具的选择和使用会被记录在 extra.tool_logs 中。

@agent.tool_func
def add(a: int, b: int) -> int:
    return a + b

response = agent.input("12+34=?").use_tool(add).get_response()
full = response.get_data(type="all")
print(full["extra"]["tool_logs"])

6) 🌐 统一模型设置 (OpenAICompatible)

一套配置即可兼容多个提供商及本地模型。

from agently import Agently

Agently.set_settings(
    "OpenAICompatible",
    {
        "base_url": "https://api.deepseek.com/v1",
        "model": "deepseek-chat",
        "auth": "DEEPSEEK_API_KEY",
    },
)

🚀 快速入门

安装

pip install -U agently

要求：Python >= 3.10，建议 Agently >= 4.0.7.2

5分钟示例

1. 结构化输出

from agently import Agently

agent = Agently.create_agent()

result = (
    agent.input("Introduce Python in one sentence and list 2 advantages")
    .output({
        "introduction": (str, "one sentence"),
        "advantages": [(str, "advantage")]
    })
    .start(ensure_keys=["introduction", "advantages[*]"])
)

print(result)

2. 工作流路由

from agently import TriggerFlow, TriggerFlowEventData

flow = TriggerFlow()

@flow.chunk
def classify_intent(data: TriggerFlowEventData):
    text = data.value
    if "price" in text:
        return "price_query"
    if "feature" in text:
        return "feature_query"
    if "buy" in text:
        return "purchase"
    return "other"

@flow.chunk
def handle_price(_: TriggerFlowEventData):
    return {"response": "Pricing depends on the plan..."}

@flow.chunk
def handle_feature(_: TriggerFlowEventData):
    return {"response": "Our product supports..."}

(
    flow.to(classify_intent)
    .match()
    .case("price_query")
    .to(handle_price)
    .case("feature_query")
    .to(handle_feature)
    .case_else()
    .to(lambda d: {"response": "What would you like to know?"})
    .end_match()
    .end()
)

print(flow.start("How much does it cost?"))

✅ 你的应用已达到专业水准了吗？