agents-best-practices

Why we included this project

Teams that run agentic workflows in Codex or Claude Code tend to discover permission drift and unplanned tool use as those workflows grow. This skill is a provider-neutral bundle that takes you from an MVP blueprint through auditing and refactoring an existing harness, without locking you into any one runtime. The core idea is that the model proposes actions and the harness validates, authorizes, executes, and records them; the same material lays out how tool routing, permissions, memory, and context compaction fit together. Because it installs as a plain skill directory, you drop it into your existing Codex or Claude Code setup instead of adopting another runtime. That makes it equally useful when you are starting from scratch or trying to rein in a harness that has already drifted.