Unpacking the Hype: A Clear Map of LLM, RAG, Agent and Agent Platforms

Agent technology is currently very popular, with new products, demos, and frameworks appearing daily. The author observes that this abundance creates a false sense of understanding, where many terms are seen but their precise roles remain unclear.

The core confusion stems from mixing concepts that belong to different layers of the AI stack. The author clarifies each term:

LLM – the foundational language and reasoning core.

RAG – the mechanism for external knowledge access.

Tool Use – the capability for external actions.

Agent – a task‑execution system that organizes these abilities around a goal.

Code Agent – a specialized form focused on software‑development scenarios.

Agent Platform – the infrastructure that hosts, manages, and schedules these components for long‑running operation.

Because these terms are not parallel concepts but belong to distinct layers—model core, augmentation modules, execution loop, and runtime platform—failing to separate them hampers learning paths. Without a layered map, learners may jump between training code, vector retrieval, and multi‑agent frameworks without forming a stable mental structure.

The author proposes a concrete learning roadmap:

Start with MiniMind to understand the minimal LLM loop, including tokenizer, pre‑training, SFT, LoRA, and DPO, focusing on the model foundation rather than performance.

Proceed to RAG from Scratch to dissect knowledge‑access components (chunking, embedding, retrieval, reranking, generation) and place them correctly in the system.

Explore Hello‑Agents to examine workflow versus agent boundaries, tool use, memory, and planning, determining when a full agent is needed versus a fixed pipeline.

Finally investigate platforms like OpenCode and OpenClaw to understand why code agents and agent platforms pose software‑engineering challenges rather than mere prompting issues.

To avoid “collection‑type” learning, the author will adopt a verification‑oriented approach: study repositories, build small projects, take structured notes, and draft an explanatory article, ensuring each stage leaves tangible artifacts.

The central question the author aims to answer is: when we talk about an “Agent,” are we referring to a LLM‑centric task system, a knowledge‑retrieval workflow, a code‑modifying development tool, or a long‑running platform that manages state and tool scheduling?

By separating these layers, the author believes future study of any related system will become clearer, allowing systematic progression from MiniMind to RAG, to Hello‑Agents, and finally to code agents and platforms.