Frontend · AI Programming · Agent

The AI Buzzword Stack, from LLM to Agent

By JacksonChen · Aug 9, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

The jump from calling a model to fielding an agent is less about model capability and more about the surrounding engineering — context, tools, memory, and guardrails. Developers who treat the model as just an API endpoint will hit a ceiling that the rest of this stack is designed to break through.

Summary

The AI vocabulary has grown faster than most developers can track, with each new term assuming fluency in the last ten. This glossary strings the high-frequency concepts into a dependency chain: from the model itself, through retrieval and tool-use, to autonomous agents and the engineering disciplines wrapping them. Every idea is grounded in the same metaphor — a very smart intern with a short memory and no hands — making the stack tangible.

It walks from LLMs, tokens, and context windows through RAG, embeddings, fine-tuning, and tool calling, then introduces MCP as a USB-C moment for AI-tool connections. Memory and Skills are framed as reusable information management, not model retraining. The Agent section distinguishes dynamic decision-making from fixed workflows, with ReAct as the canonical think-act-observe loop.

The final third covers the engineering roles emerging around models: Prompt Engineering (what to ask), Context Engineering (what information to surround the model with), Harness Engineering (the runtime framework including sandboxing, permissions, and retries), and Loop Engineering (stopping infinite token burn). AGI is the horizon, but the immediate work is designing the environment the model runs in.

Takeaways

— An LLM predicts the next token; its apparent reasoning, coding, and planning are emergent from that single task.

— Tokens are the model's unit of work — API billing, context length, and throughput all count in tokens, not words.

— A context window is the total information a model can see in one interaction; old conversation falls out, it isn't "forgotten."

— RAG gives the model access to private or recent data by retrieving relevant chunks from a vector store before answering.

— Embedding converts text into numerical vectors so semantic similarity can be computed; production RAG usually combines vector and keyword search.

— Fine-tuning changes how a model behaves; RAG changes what it knows. They solve different problems.

— Tool calling lets the model decide to invoke an external function, but the host application executes it — the model only issues the request.

— MCP aims to standardize how AI applications connect to tools, analogous to USB-C replacing proprietary connectors.

— Memory externalizes important context across sessions so an agent can resume work without replaying the entire history.

— A Skill is a reusable task specification — steps, tools, output format — loaded at runtime without retraining.

— An Agent differs from a workflow when the model dynamically decides the next step rather than following a pre-scripted path.

— ReAct (Reason + Act) is the core agent loop: think, act, observe, then think again.

— Prompt Engineering optimizes the question; Context Engineering designs the entire information environment the model operates in.

— Harness Engineering wraps the model in a runtime with sandboxing, permissions, retries, and tool wiring — the model is the engine, the harness is the car.

— Loop Engineering tackles when to stop an agent cycle, how to compress context, and when to escalate to a human.

Conclusions

The glossary's structure — model → enhancement → agent → engineering → AGI — itself encodes a thesis: raw model intelligence is necessary but insufficient; the engineering layers above it are where production behavior is determined.

Framing Memory and Skills as external information management, not model internals, clarifies why agent reliability depends more on retrieval and prompt assembly than on the base model's recall.

The distinction between Prompt Engineering and Context Engineering marks a genuine shift in practitioner focus — from crafting a single instruction to curating the model's entire runtime information supply.

Harness Engineering is the least discussed but highest-leverage concept here; it's the difference between a demo that works once and an agent that runs unattended without trashing a production database.

Concepts & terms

Token

The atomic unit of text that a language model processes. Text is split into tokens before being fed to the model; API costs, context limits, and throughput are all measured in tokens.

Context Window

The maximum number of tokens a model can consider in a single interaction, spanning input, conversation history, system prompts, tool definitions, and output. Content beyond this window is not remembered — it was never seen.

RAG (Retrieval-Augmented Generation)

A pattern that retrieves relevant documents from an external knowledge base and supplies them to the model at inference time, letting it answer with data it wasn't trained on.

Embedding

A numerical vector representation of text that captures semantic meaning, enabling similarity search. Production systems often combine embedding-based vector search with traditional keyword matching for reliability.

Fine-tuning

Further training a pre-trained model on a specific dataset to alter its behavior, style, or domain expertise — distinct from RAG, which changes what the model knows without changing how it behaves.

Tool Calling

A mechanism where the model outputs a structured request to invoke an external function (API, database, code execution), but the host application performs the actual execution and returns the result.

MCP (Model Context Protocol)

A proposed standard protocol for connecting AI applications to external tools, aiming to replace bespoke per-tool integrations with a uniform interface — analogous to USB-C for peripherals.

Agent

An AI system that autonomously pursues a goal by dynamically deciding its next action: analyzing the task, planning, calling tools, observing results, and adjusting — as opposed to following a fixed workflow.

ReAct

A reasoning-and-acting loop pattern for agents: Reason about the situation, take an Action, Observe the result, then reason again — the canonical think-act-observe cycle.

Harness Engineering

The practice of building the runtime framework around a model — system prompts, tool wiring, sandboxing, permissions, retries, and context management — that determines whether it runs safely and stably in production.

Loop Engineering

The discipline of controlling an agent's iterative cycle: setting stop conditions, preventing infinite token consumption, deciding when to compress context or escalate to a human.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗