Claude Code多智能体自动化测试：Pytest虚拟测试团队实战_CG群

本课程教你用Claude Code构建多智能体虚拟测试团队，编写6个专职子智能体，串联规划、编写、评审与调查，实现Pytest与Playwright自动化测试闭环。

原始标题：Agentic Automation — A Pytest Framework with Claude Code

Agentic Automation — A Pytest Framework with Claude Code

这门课程教学员如何使用 Claude Code 摆脱传统的手写测试脚本，转而构建一个由 AI 智能体（Agents）组成的“虚拟测试团队”。你将学习从零编写 6 个各司其职的专业子智能体（涵盖架构规划、代码编写、代码评审、失败原因调查等），并将它们串联成一个自动化、能互相交接工作的多智能体工程闭环。

在技术实现上，课程聚焦于 Python、Pytest 和 Playwright 框架，重点教授如何利用“规划模式”分离思考与执行，以及如何在严格的安全护栏与权限策略下优化智能体。这适合有一定 Python 基础、希望将前沿的 AI 智能体技术应用到软件测试自动化中的中级开发者。

Agentic Automation: Skills, Subagents, Planning, and Test Automation

What you’ll learn
⚡ Design and author purpose-built Claude Code subagents — giving each a narrow mandate and exactly the tools it needs
⚡ Separate thinking from doing with plan mode — produce a reviewable plan before writing code
⚡ Orchestrate a multi-agent engineering loop — plan → build → review → test — where a lean orchestrator routes work across agents
⚡ Turn requirements into a running Pytest + Playwright framework — mapping user stories and acceptance criteria into non-flaky automated tests
⚡ Work within enforceable guardrails and improve agents from real results — building inside allow/deny permission policies

Requirements
❗ Comfortable with Python, Pytest, and Git; some Playwright and Claude Code exposure helps

Description

Agentic Automation Course

Build a test-automation framework where the work is done by AI agents, not scripts you run by hand.
Most automation courses teach you to write scripts. This one teaches you to build ateam — a roster of purpose-built Claude Code subagents that plan, write, review, run, and debug tests by reasoning about the work, using their own tools, and handing off to each other in a real engineering loop.

You’ll author each specialist from scratch and wire it into a working pipeline

✅plan-architect researches a task and returns a reviewable plan before a line of code is written

✅automation-writer turns user stories and acceptance criteria into non-flaky Pytest + Playwright tests

✅senior-code-reviewer reviews the diff read-only and opens the PR on approval

✅test-suite-planner decomposes a story into prioritised, independent suites as Jira sub-tasks

✅test-runner executes each reviewed suite and posts a sanitized verdict

✅qa-failure-investigator tells a test bug from a product bug — byattempting the fix — and files evidence-backed defects

A leanorchestrator routes work between them and carries context across the hand-offs. Because the heavy lifting stays inside each agent’s isolated context window, the orchestrator runs a long, multi-stage workflow without exhausting its own — in one measured build it held at~5%

context used while ~250K tokens of work happened inside the subagents.

What makes it different
✅“Agentic” is literal. Each agent is an autonomous LLM worker with a narrow mandate, its own context, and exactly the tools it needs — toolsare the capability boundary, so a read-only reviewer physically can’t mutate the repo.

✅Dogfooded throughout. Every agent is exercised on real work in the course’s own repo, driving a live Jira board and a real browser against a running app. Whatever breaks gets fed back into that agent’s prompt — so you learn to improve agents the way you’ll actually improve them.

✅Guardrails you workwith . Destructive, staging, commit, push, and PR-create commands are denied outright; sanctioned actions are wrapped in allow-listed skills. You learn to build inside enforceable policy, not around it.

The pipelineplan → build → review → test, with a defect fork when a suite catches a real product bug. Progress is visible on the Jira board through a single parent-story status, one owner per transition

To Do → In Planning → Ready for Dev → In Progress → In Review → In Testing → Done

You’ll practice
Matching tools to intent · planning before doing · delegating execution to keep an orchestrator’s context lean · separating “it compiles” from verified and “review” from security audit · and feeding real findings back into agents so each dogfood failure pre-empts its whole class

of bug next time.

Stack: Python · Pytest · Playwright · pipenv · pytest-html + Allure · custom Claude Code subagents and skills · live Jira Cloud.

Educational course material — designed to be worked through interactively with Claude Code against a sandbox Jira project and a disposable app-under-test.

Who this course is for
⭐ Developers and automation engineers learning to build with Claude Code subagents

隐藏内容

此处内容需要权限查看