Summary

This repository provides a provider-neutral framework for designing, auditing, and generating blueprints for agentic harnesses. It targets engineers building AI agents, offering a disciplined approach to integrating models with real systems by emphasizing a robust harness that validates, authorizes, and executes model-proposed actions, enhancing safety and auditability.

How It Works

The core principle is "the harness acts, not the model." It implements a model-tool-observation loop where the model proposes actions, and application code validates, authorizes, executes, and records outcomes. Key features include narrow, typed tools, runtime permission checks, budgets for long-running tasks (step, time, cost), and context/memory compaction preserving active state. Risk stratification dictates distinct permission paths for operations, with high-risk actions requiring external approval.

Quick Start & Requirements

Install globally via npx skills add DenisSergeevitch/agents-best-practices -g or manually clone into agent runtime directories (e.g., ~/.codex/skills/). Prerequisites include Node.js/npm (for npx) and git. Compatible with agent runtimes like Codex and Claude Code, supporting provider-neutral API patterns.

Highlighted Details

• Provider-neutral design for agent harness architecture.
• Runtime discipline: "The harness acts, not the model."
• Extensive reference guides cover MVP blueprints, agentic loops, tool/permission design, context management, prompt caching, and security evaluations.
• Risk-based tool exposure and execution, requiring external