Agent Knowledge Base
An AI Agent knowledge system built for web frontend developers: from your first LLM API call to complete, interview-defensible practical skills. Everything is aligned with what 2026 job postings actually demand — "AI Application Engineer / Agent Engineer" roles don't ask you to train models; they ask you to engineer LLM capability into real business: call APIs well, understand RAG, orchestrate agents, build streaming UIs, and reason about evaluation and cost.
First, the mindset: what this market is actually hiring for
Translating JD buzzwords into plain requirements:
| JD wording | What they really mean | Covered in |
|---|---|---|
| "Familiar with LLM internals" | Understand the engineering meaning of tokens / context windows / sampling params — no backprop by hand | LLM Basics & APIs |
| "Strong prompt writing & tuning" | Structured prompts, few-shot, structured output, bad-case iteration | Prompt Engineering |
| "Task planning, tool use, memory, autonomous decisions" | The agent-loop quartet; implement a minimal agent from scratch | Agent Core Principles |
| "Deep understanding of RAG (reranking, multi-way recall, KG integration)" | Every pipeline stage: why, what to pick, how to tune | RAG |
| "Understand MCP; able to build custom MCP servers" | Know the protocol + have shipped a TS SDK server | Tool Ecosystem & MCP |
| "Familiar with LangChain / LangGraph / Dify" | Know each framework's niche and trade-offs, not just doc-copying | Frameworks & Selection |
| "Streaming output, multi-platform delivery" | SSE parsing, streaming Markdown rendering, typewriter & abort | Frontend Agent Engineering |
| "Agent evaluation & production deployment" | Eval datasets, LLM-as-judge, observability, cost control | Production & Career Guide |
Know your edge as a frontend dev: TypeScript is now a first-class citizen of the agent ecosystem — Vercel AI SDK, LangGraph.js, the official MCP TypeScript SDK, and the Claude Agent SDK all ship full TS support. The interaction layer of AI products (streaming UIs, workflow visualizers, low-code orchestration) is literally the frontend's home turf. Don't fight algorithm engineers on model internals — build your advantage in product delivery and agent engineering.
Knowledge map & page navigation
Read in table order — the first four pages are foundations, the last four are practice:
| # | Page | What you gain |
|---|---|---|
| 1 | LLM Basics & API Calling | Tokens/context/sampling mental model, OpenAI-compatible API, SSE streaming, model & pricing landscape |
| 2 | Prompt Engineering & Structured Output | Prompt templates, few-shot/CoT, JSON mode, the full Function Calling mechanism |
| 3 | Agent Core Principles | A 50-line hand-built agent loop, ReAct, planning/tools/memory/reflection, workflow vs agent |
| 4 | RAG | End-to-end engineering judgment: chunking → embeddings → vector DB → hybrid recall → rerank → eval |
| 5 | Tool Ecosystem & MCP | MCP architecture and primitives; write your own MCP server with the TS SDK |
| 6 | Frameworks & Selection | Where LangGraph.js / Vercel AI SDK / OpenAI Agents SDK / Mastra / Dify·Coze each fit |
| 7 | Frontend Agent Engineering | Full streaming chat stack: SSE parsing, useChat, Markdown rendering, tool-progress UI, XSS defense |
| 8 | Production & Career Guide | Eval, observability, cost & latency optimization, safety, plus resume projects and top interview questions |
Recommended practice track (three resume-worthy projects)
Interviewers distinguish "followed a tutorial" from "built a project" brutally — can you explain exception handling and selection rationale? Build in this order:
- Streaming AI chat app (1 week): Vercel AI SDK or raw SSE + your own UI. Must-haves: typewriter rendering, stop generation (AbortController), multi-turn context, streaming Markdown parsing. → Covers pages 1, 2, 7.
- Document-QA RAG assistant (2 weeks): upload PDF → chunk → embed (pgvector/sqlite-vec) → hybrid recall + rerank → citation source view. Must-haves: chunking strategy A/B experiments, retrieval quality metrics. → Covers page 4; the project you'll be grilled on most.
- Tool-calling agent (2–3 weeks): a multi-step agent that checks weather / searches the web / reads files, or write an MCP server plugged into Claude/Cursor. Must-haves: tool-failure retry, max-iteration guardrails, trace logs. → Covers pages 3, 5, 6.
Learning resources
- OpenAI Platform Docs — the de facto API standard; understand the compatible protocol here
- Anthropic Docs and Building Effective Agents — the canonical article on agent design patterns (source of the workflow-vs-agent taxonomy)
- Vercel AI SDK / LangGraph.js / MCP official site — the three TS-ecosystem staples
- LangChain / LlamaIndex / Langfuse — Python orchestration, RAG-focused framework, open-source observability
- Dify / Coze — low-code platforms for fast idea validation
- DeepLearning.AI short courses — Andrew Ng's series, 1–2 hours each, excellent RAG/agent coverage
Ready? Start with lesson one: LLM Basics & API Calling →