Building Agent Systems

What it does

A practical engineering reference for building LLM-powered agents and multi-agent systems. Covers single-agent patterns (ReAct, plan-execute, reflection), multi-agent coordination (team creation, file locking, role decomposition), LLM security (prompt-injection defenses, jailbreak mitigation, output filtering), and RAG architecture (chunking, hybrid retrieval, reranking). Includes operational guidance: scale decisions, lifecycle rules, and integration points to provisioning and CI.

When to use it

Use this when designing or hardening production agent systems: choosing single vs multi-agent architectures, implementing RAG pipelines, establishing prompt/version control, or defining security/evaluation guardrails for deployed LLM apps. Helpful for architects, SREs, and lead engineers building agent orchestration.

What's included

• Scripts: none bundled in this skill (has_references=true)
• References: extensive reference docs for agent-dev, multi-agent coordination, protocol design, LLM security, and RAG systems (referenced under references/).
• Instructions: scale-decision heuristics, multi-agent enablement criteria, file-locking and downgrade paths, and explicit rules for prompt versioning and I/O validation.

Compatible agents

Best suited for agent frameworks and tools that support composition and tool-calling (Claude Code, Codex-style agents, Gemini CLI, and OpenClaw orchestrations).