A1 · Augmented LLM
Family: Building blocks
A single model call, enriched with documents, tools and memory.
Principle
Besides the question, the model gets retrieved passages, tools it can call and a memory. It writes its own queries, picks its tools and decides what to keep. This is the building block of every other pattern.
The problem it solves
A model on its own knows neither your documents nor today’s data, and cannot act on anything. Yet many needs can be met without an agent.
When to use it
- The task fits in one pass: classification, extraction, summarisation, translation, question answering over documents.
- You are starting a project: a well-built single call is the baseline to beat before adopting any heavier pattern.
When to avoid it
- The number of steps depends on what is found along the way: move to the agent loop (C1).
- The task breaks down into fixed steps you want to control one by one: prefer prompt chaining (B1).
Common pitfalls
The anti-patterns that misuse this pattern, best established first. Each sheet describes the symptoms, the public incidents and the remedies.
Framing and choice of level
- AP05 · One model for everything — The same model serves every step, from trivial reformatting to hard reasoning.
Instructions
- AP21 · Agent instruction bloat — An instructions file that grows with every incident, until important rules get ignored.
- AP22 · Contradictory instructions — Two incompatible instructions in the same prompt, and the model spends its reasoning reconciling them instead of acting.
- AP23 · Business logic hardcoded in the prompt — A tree of “if… then” written in prose, which a probabilistic model executes poorly and which grows with every bug.
- AP24 · Legacy over-prompting — “CRITICAL: You MUST” lines written for the old model, which the new one takes literally.
Context and memory
- AP34 · Reflexive vector RAG for code — A vector database to find code, where `grep` and reading files do better and more simply.
Security and permissions
- AP60 · Uncontrolled rendering as an exit channel — The interface loads on its own the images whose URL the model writes: an exit channel nobody designed.
Oversight, operations and governance
- AP81 · Hard vendor coupling — The agent rests on a stateful API or a specific model with no isolation layer: a forced migration becomes a rewrite.
Synonyms by vendor
- Augmented LLM (Anthropic)
- L1 Connected Problem-Solver (Google (Introduction to Agents))
- Knowledge retrieval / RAG (A. Gulli (ch. 14))
- Memory-augmented agents (AWS)
Neighbouring patterns
Sources
- AI agent orchestration patterns — Microsoft (Azure Architecture Center), 2026-02-12
- Building effective agents — Anthropic (E. Schluntz, B. Zhang), 2024-12-19
Content reviewed on 2026-10-02