agents-best-practices

agents-best-practices is a provider-neutral Agent Skill, a SKILL.md with reference documents, that helps an AI assistant design, audit, refactor and explain agentic harnesses. Its core idea is stated in the opening quote: the model proposes actions, and the harness validates, authorizes, executes, records and returns observations. The README says it applies beyond coding agents, to research, support, operations, finance, data analysis and other workflow agents.

The skill covers a model-tool-observation loop, narrow typed tools with structured results, permission checks outside the model, planning mode and approval-gated execution, goal loops with budgets and stop rules, context and memory, skills and MCP connector governance, prompt-cache-aware layout, observability, evals and launch gates. The README shows three use cases: generating an MVP blueprint (its example is an approval-gated renewal-risk agent with a minimal tool list and a launch gate of 20 historical accounts and 80% human acceptance), auditing an existing harness, and designing tools and permissions. It is explicit about what it is not: not only for coding agents, not a multi-agent framework by default, not a replacement for runtime authorization or sandboxing. Install with `npx skills add DenisSergeevitch/agents-best-practices -g` or clone it into the Codex or Claude Code skills directory. It is MIT licensed.