AI coding agents: harnesses, MCP and code review
A coding agent is a model running in a loop with tools. Once you understand the loop, the harness around it and the protocol it uses to reach your tools, you can tell when an agent will do good work and when it will confidently waste your afternoon.
このページの本文は英語です。
コースを始める
レッスン
What a coding agent actually is
The model, the loop, the tools and the context window, and why the same model behaves differently in different products.
Inside the harness: context, tools and permissions
Instruction files, compaction, subagents, hooks, permission modes and the sandbox, and which of them the model is free to ignore.
MCP: how agents connect to tools and data
Hosts, clients and servers; tools, resources and prompts; the stateless 2026-07-28 protocol, its transports and authorization.
Build a small MCP server
Write a working server with the official SDK, test it with the Inspector and connect it to an agent.
Automated code review with Claude
How Claude reviews pull requests, what it catches, what it misses and how to configure it for your repository.
A workflow that gets good results from agents
Define done first, plan before editing, keep diffs small, let checks close the loop and read the diff before you trust it.
Agent safety: prompt injection, secrets and permissions
What can go wrong when an agent reads untrusted text or holds your credentials, and the settings that limit the damage.
Interview Prep パス · プランを比較