# CPU Bytes > Clear, citation-ready explainers on distributed systems, LLM architecture, and backend performance. CPU Bytes is an engineering knowledge platform built for developers and AI copilots. Every article is structured with clear headings, direct answers, and structured data so AI systems can cite content accurately. ## Topics - [AI Agents & Agentic Systems](https://cpu-bytes.com/blogs/ai-agentic-systems): What AI agents are and how to build them: autonomous AI agents, agent architecture, multi-agent systems, and agentic workflows. - [What's Actually Inside an AI Agent](https://cpu-bytes.com/blogs/ai-agentic-systems/ai-agents-as-systems): A layer-by-layer look at what's inside an AI agent — the tool-calling loop, memory, and the production infrastructure that keeps it alive. - [Tokens](https://cpu-bytes.com/blogs/ai-agentic-systems/tokens): How tokenization works, token limits, context windows, and their impact on AI agent behavior. - [RAG (Retrieval-Augmented Generation)](https://cpu-bytes.com/blogs/ai-agentic-systems/rag): How retrieval, embeddings, and generation combine to let an LLM answer questions about your private data without retraining it. - [CAG (Cache-Augmented Generation)](https://cpu-bytes.com/blogs/ai-agentic-systems/cag): Prefilling a knowledge base once and reusing its KV cache across questions — how cache-augmented generation works and when it beats RAG. - [CUDA](https://cpu-bytes.com/blogs/cuda): GPU programming with CUDA — kernels, threads, and blocks, and why deep CUDA usage creates real vendor lock-in against AMD and Intel alternatives. - [What Is CUDA, and Why Does GPU Programming Create Vendor Lock-In?](https://cpu-bytes.com/blogs/cuda/what-is-cuda): CUDA isn't a language, it's NVIDIA's whole GPU computing platform. Here's what a kernel launch actually does, how threads and blocks work, and why deep CUDA usage becomes so hard to leave. - [Field Notes](https://cpu-bytes.com/blogs/field-notes): Short-form takes on what's actually happening in AI right now — new releases, changed rules, and the things worth noticing as they happen. - [AI Watermarking Just Changed the Rules, and Open Models Make It Messy](https://cpu-bytes.com/blogs/field-notes/ai-watermarking-and-open-models): Anthropic now embeds an invisible watermark in Claude's output at generation time. Here's how that actually works, what it means for SEO, and why open-weight models can't be forced to carry the same signal. ## Articles - [AI Watermarking Just Changed the Rules, and Open Models Make It Messy](https://cpu-bytes.com/blogs/field-notes/ai-watermarking-and-open-models): Anthropic now embeds an invisible watermark in Claude's output at generation time, hidden in word choice rather than metadata. Here's how it works, what it means for AI content and SEO, and why open-weight models can't be forced to carry the same signal. - [What's Actually Inside an AI Agent](https://cpu-bytes.com/blogs/ai-agentic-systems/ai-agents-as-systems): An AI agent is a tool-calling while loop until real users show up. What's inside one: the agent loop, tool calls, memory, tracing, evals, and guardrails. - [What Is CUDA, and Why Does GPU Programming Create Vendor Lock-In?](https://cpu-bytes.com/blogs/cuda/what-is-cuda): What CUDA actually is, how kernels, blocks and threads work, and why deep CUDA usage creates real vendor lock-in against AMD ROCm and Intel oneAPI. - [Building a Cache-Augmented Generation System From Scratch](https://cpu-bytes.com/blogs/ai-agentic-systems/cag): How cache-augmented generation actually works, built from scratch with PyTorch and Hugging Face Transformers by prefilling a KV cache once and reusing it across questions. - [What RAG Actually Is (and Why Your LLM Doesn't Know Your Company Data)](https://cpu-bytes.com/blogs/ai-agentic-systems/rag): How does RAG actually work? It doesn't retrain your LLM, it changes what the LLM sees at query time by combining indexing, embeddings, vector search, and generation over your private data. - [Tokens! What They Actually Are](https://cpu-bytes.com/blogs/ai-agentic-systems/tokens): What are tokens in AI, really? Not words — they're subword units a tokenizer learns from raw text. Here's how tokenization, BPE, and token IDs actually work, and why it affects cost. ## Contact - Website: https://cpu-bytes.com - Email: contact@cpu-bytes.com