An opinionated ops manual for LLM agent systems

What it does This repository is a curated handbook for building, evaluating, and shipping modern LLM agent systems. It covers the entire stack—identity, memory, MCP integrations, skills, safety, observability, and provider strategy—and bundles conceptual guides with runnable Python utilities like an LLMProvider abstraction and a fallback router.

The interesting bit The project is designed to be consumed by coding agents themselves: it ships llms.txt, tool-agnostic prompts for Claude Code, Cursor, Codex, and others, and a machine-readable DESIGN.md spec. It is essentially a textbook that practices what it preaches.

Key highlights

• 24+ provider adapters with a task-class router that chains fallbacks across Groq, DeepSeek, Together, and OpenRouter—not just docs, but working code.
• 100+ curated agent skeletons, framework comparisons, evaluation tools, and RAG tutorials preserved in one place.
• Full-stack conceptual coverage: prompt-injection defense, instruction hierarchy, memory taxonomy, skill maturity models, and production readiness checklists.
• Tool-neutral coding-agent workflows: repo audit prompts, modernization templates, and task contracts that work across Claude Code, Cursor, Aider, Cline, and Codex.
• Design doc discipline: ADR guides, design reviews, and a machine-readable DESIGN.md spec for design tokens.

Verdict Reach for this if you are architecting or shipping LLM agents and want a production-oriented map of the territory with starter code. Look elsewhere if you need a single, monolithic framework to install; this is a reference library and operating manual.