The Augmented LLM

The foundational building block of every agentic system: an LLM enhanced with augmentations — retrieval, tools, and memory — that the model actively drives itself (generating its own search queries, selecting tools, deciding what to retain). Every workflow and agent is composed from this unit; the rest of the design is how you wire augmented calls together.

Two implementation priorities: tailor the augmentations to your specific use case, and give the model an easy, well-documented interface to them. The Model Context Protocol (MCP) is one way to expose augmentations, letting a simple client integrate a growing ecosystem of third-party tools — but it is a means, not a requirement.

Claims

The augmented LLM (LLM + retrieval + tools + memory) is the basic building block of all agentic systems. principle — durable: every higher pattern reduces to composing augmented calls, so this is the stable primitive.
Modern models can actively drive their own augmentations — generate queries, pick tools, choose what to retain. observation — a capability note about current models that makes the building block viable.
Tailor augmentations to the use case and expose them through an easy, well-documented interface. best practice — context: implementing any augmented call; the payoff of a clean interface is contingent on the model actually having to use it well, which links directly to Agent-Computer Interface (ACI).
**The composition is RAG (information) + tools (capability extension) + planning (complex tasks)
- memory (continuity); agent tools fall into knowledge-augmentation, capability-extension, and write-action categories, with write actions requiring the strongest safeguards.** observation — the AI Engineering textbook builds the same augmented unit from its component list, and adds the tool taxonomy; independent corroboration of the building-block framing.
Agent success rate decays with each step because errors compound, and the stakes rise because tools are powerful — so agents need more capable models than single-shot apps. principle — the cost of composing many augmented calls into a multi-step loop; names why agentic reliability is hard.

Workflows vs Agents — both are built by composing augmented LLM calls.
Agentic Workflow Patterns — the patterns assume each call is an augmented LLM.
Agent-Computer Interface (ACI) — "well-documented interface for the LLM" is the ACI discipline in miniature.
Adaptation Strategy Ladder — the augmented LLM is what the RAG and tool rungs of the ladder build toward.
Layered Agent Memory — the memory augmentation of this unit, tiered.
Distillate: Building Effective Agents.
Distillate: AI Engineering in 76 Minutes — Chip Huyen's Book, Speedrun — textbook corroboration of the RAG+tools+memory+planning composition, plus the agent-tool taxonomy and compounding-error note.

The Augmented LLM

Claims

Related

Linked from