What is agents-best-practices?

Agents Best Practices is an Agent Skill, a document plus reference files that AI coding assistants like Codex and Claude Code can load, that teaches how to design safe and reliable agent systems. It is not tied to one AI provider and applies beyond coding assistants, covering agents for research, customer support, operations, sales, finance, data analysis, procurement, legal work, healthcare, and education. The core idea is that an AI model should only propose actions, while a separate runtime checks, authorizes, executes, records, and returns the results of those actions. The skill teaches you how to build that runtime: a typed loop where the model calls a tool, the tool call is validated and permission checked, then either executed or paused for human approval, with every step producing a structured observation the model can read back. Once installed, the skill activates automatically whenever a conversation touches agent architecture, tool permissions, planning mode, memory and context compaction, connectors, observability, evaluation, prompt caching, or getting an agent ready for production. It can generate a minimal but production safe blueprint for a new agent, audit an existing agent that behaves unpredictably or loses track of its state, or help design tools and permission boundaries so an agent does not get overly broad access to systems like Slack, Linear, or an internal deploy API. The project is organized as a main skill file plus a set of reference documents covering topics such as the agent loop, context and memory compaction, tool and permission design, skills and connector governance, prompt caching and cost, provider specific API patterns, and pre launch checklists. You install it either with the skills command line tool, by asking your AI agent to clone it into its skills folder, or by cloning it manually into the correct directory for Codex or Claude Code. The project is released under the MIT license.