agent-architectures

一个可直接使用的 Agent Skill。涵盖六大生物启发模式、六大设计原则，以及可落地的操作手册。

Cognitive Outsourcing (CO): edge-AI architecture empowering 0.8B–4B models with dynamic access to cloud LLM teachers, tools, and local caches via Suspend-and-Inject Generation (SIG). Keeps KV-cache intact across tool calls—up to 96% prefill savings and 1.57× speedup.

An interactive benchmarking dashboard to compare and visualize multi-agent architectures (single, hybrid, dynamic swarm) in real-time. Analyze tradeoffs in quality, latency, and cost. Inspired by Google's "Science of Scaling Agent Systems" paper and powered by LangGraph + Gemini.

Reactive, BDI, layered, and utility-based agent design patterns with runnable examples.

Classical AI agent architectures are not obsolete: an LLM upgrades one component inside them and leaves the architecture standing. Twenty runnable before/after pairs, nine config-driven agents across all five architectures, and every claim shipping with the command that checks it.

A collection of the messy problems I've hit, the weird fixes I found, and the refactors that finally made things behave. Basically, my debugging diary, but useful.

Deterministic companion demo for bounded mirror-recognition evaluation under explicit governance.

AI learning and experimentation repo — structured study notes, code experiments, and research on RAG systems, agent architectures, and LLM evaluation methods.

AI Agent Reference Architecture Patterns

Visual Streamlit playground implementing six LLM agent architectures (Single LLM, Tool Calling, ReAct, Plan-and-Execute, Reflection, Reflexion) on a shared event-driven runtime — showing how each plans, calls tools, and self-corrects.