agent-architecture

This skill provides a robust architectural framework for building sophisticated, tool-using AI agents designed for production environments. It is intended for software engineers and AI architects who need to move beyond simple chat interfaces to implement stateful, multi-agent orchestrations. By leveraging Anthropic's Claude models, LangGraph for graph-based workflow control, and Vercel AI SDK v6 for efficient streaming, this skill ensures that agentic logic remains maintainable, predictable, and cost-effective.

• Multi-agent orchestration using LangGraph for complex, stateful reasoning tasks.
• Implementation of mandatory guardrails, including MAX_TURNS and MAX_TOKENS_PER_RUN, to prevent runaway costs and infinite loops.
• Episodic memory management utilizing pgvector for persistent context, retrieval-augmented generation (RAG), and state handling.
• Full-stack integration patterns including SSE streaming from FastAPI backends to Next.js frontends.
• Model Context Protocol (MCP) overview for standardized tool connectivity.
• Failure mode analysis with curated anti-pattern and fix code pairs to help developers identify and remediate common agentic pitfalls.
• Seamless transition from basic one-call agent logic to complex, autonomous loop-based systems.

When utilizing this skill, expect to provide natural language requirements or architectural designs; the skill outputs verified implementation templates, configuration snippets, and best-practice workflows. Typical inputs involve defining agent objectives, tool signatures, or state schema requirements. Expected outputs include modular Python or TypeScript code, safety configuration constants, and infrastructure patterns for persistent memory. Users should ensure their development environment has access to necessary LLM provider keys and that database dependencies like PostgreSQL/pgvector are prepared for memory-intensive operations. Always prioritize the manual loop implementation provided in the documentation to ensure total control over guardrail placement, especially in cost-sensitive production deployments.