在以下情况下不要使用本技能：

• MCP 服务器注册或 MCP 传输选择
• 模型设置、输出模式设计或结构化流式传输作为主要问题
• TriggerFlow 工作流设计
• FastAPI 服务暴露

工作流程

1. 在选择工具应如何注册或附加到代理时，请从 references/registration-and-selection.md 开始。
2. 当问题涉及循环行为、轮次、并发、超时或自定义处理器时，请阅读 references/tool-loop-and-control.md。
3. 当任务需要使用 Search、Browse 或 Cmd 时，请阅读 references/builtin-tools.md。
4. 当任务涉及工具日志、操作结果或响应检查时，请阅读 references/observability-and-results.md。
5. 如果工具来自 MCP 服务器，请切换到 agently-mcp。
6. 如果任务变为工作流或多步骤编排问题，请切换到 agently-triggerflow-playbook。
7. 如果行为仍然看起来不正确，请使用 references/troubleshooting.md。

核心心智模型

Agently 工具是请求时的能力扩展。

• 工具在共享的工具管理器上注册
• agent.use_tools(...) 通过标签将选定的工具限定到一个代理的作用域内
• 工具循环进行计划、执行，并将工具结果反馈到同一请求中
• 最终响应在元数据中保留工具日志

Agently 对于工具工作的指导应保持异步优先：

• 优先使用异步工具函数
• 优先使用异步内置工具调用
• 明确限制工具循环轮次和并发数

选择规则

• 本地确定性 Python 辅助函数 -> @agent.tool_func 或 register_tool(...)
• 将一个或多个已注册的工具附加到代理 -> use_tools(...)
• 使用现有辅助程序暴露网络搜索、浏览或 shell 访问 -> 内置工具
• 调整最大轮次、并发数或超时 -> set_tool_loop(...)
• 替换计划或执行策略 -> 自定义计划分析或执行处理器
• 外部工具服务器而非本地 Python 函数 -> agently-mcp

重要边界

• 工具循环是代理/请求的能力，不是 TriggerFlow 的替代品
• MCP 注册属于 agently-mcp，即使 MCP 工具稍后参与相同的工具循环
• 响应元数据和输出控制是独立关注点；当主要问题是响应消费时，请使用 agently-output-control

参考资料

• references/source-map.md
• references/registration-and-selection.md
• references/tool-loop-and-control.md
• references/builtin-tools.md
• references/observability-and-results.md
• references/troubleshooting.md

每周安装数

1

仓库

agentera/agently-skills

首次出现

1 天前

安全审计

Gen Agent Trust HubPassSocketPassSnykWarn

安装于

mcpjam1

claude-code1

replit1

junie1

windsurf1

zencoder1

Do not use this skill for:

• MCP server registration or MCP transport choices
• model setup, output schema design, or structured streaming as the main problem
• TriggerFlow workflow design
• FastAPI service exposure

Workflow

1. Start with references/registration-and-selection.md when choosing how a tool should be registered or attached to an agent.
2. Read references/tool-loop-and-control.md when the issue is loop behavior, rounds, concurrency, timeout, or custom handlers.
3. Read references/builtin-tools.md when the task should use Search, Browse, or Cmd.
4. Read references/observability-and-results.md when the task is about tool logs, action results, or response inspection.
5. If the tool comes from an MCP server, switch to agently-mcp.
6. If the task becomes a workflow or multi-step orchestration problem, switch to agently-triggerflow-playbook.
7. If behavior still looks wrong, use references/troubleshooting.md.

Core Mental Model

Agently tools are request-time capability extensions.

• tools are registered on the shared tool manager
• agent.use_tools(...) scopes selected tools to one agent by tag
• the tool loop plans, executes, and feeds tool results back into the same request
• the final response keeps tool logs in metadata

Agently guidance for tool work should remain async-first:

• prefer async tool functions
• prefer async built-in tool calls
• cap tool-loop rounds and concurrency explicitly

Selection Rules

• local deterministic Python helper -> @agent.tool_func or register_tool(...)
• attach one or more already-registered tools to an agent -> use_tools(...)
• expose web search, browsing, or shell access with existing helpers -> built-in tools
• tune max rounds, concurrency, or timeout -> set_tool_loop(...)
• replace planning or execution policy -> custom plan-analysis or execution handler
• external tool server instead of local Python function -> agently-mcp

Important Boundaries

• the tool loop is an agent/request capability, not a TriggerFlow substitute
• MCP registration belongs to agently-mcp, even though MCP tools later participate in the same tool loop
• response metadata and output control are separate concerns; use agently-output-control when the main problem is response consumption

References

• references/source-map.md
• references/registration-and-selection.md
• references/tool-loop-and-control.md
• references/builtin-tools.md
• references/observability-and-results.md
• references/troubleshooting.md

Weekly Installs

1

Repository

agentera/agently-skills

First Seen

1 day ago

Security Audits

Gen Agent Trust HubPassSocketPassSnykWarn

Installed on

mcpjam1

claude-code1

replit1

junie1

windsurf1

zencoder1