What is an augmented LLM?
An augmented LLM is a language model extended with retrieval, tools and memory, so it can fetch information, take actions and keep state; Anthropic names it the basic building block of agentic systems.
The term comes from Anthropic's "Building effective agents" (December 2024). Before describing any workflow or agent pattern, the post sets out one building block: a model that can generate its own search queries, choose which tools to call and decide what information to keep. Every pattern after it, from prompt chaining to fully autonomous agents, is built from one or more of these blocks.
The point of naming it is to separate two questions. First, does each model call have the right capabilities and a clean interface to them? Second, how are those calls arranged? Teams that skip the first question end up debugging orchestration when the real problem is a vague tool description.
The three augmentations
Retrieval
The model can search documents, records or the web and pull relevant passages into context, instead of relying on what it learned in training.
Tools
The model can call functions that read or change external systems: look up an account, create a ticket, run code.
Memory
The model can write down and later read back information that outlives one request, such as user preferences or notes from earlier sessions.
Designing the interface
Anthropic's advice in the same post is to tailor the augmentations to the use case and give the model an easy, well-documented interface to each one. It names the Model Context Protocol as one way to connect third-party tools.
Anthropic's later "Writing effective tools for agents" (September 2025) makes this concrete: a few tools that match real workflows rather than one wrapper per API endpoint, namespaced names, responses that return meaningful context in a compact form, and error messages written to steer the model back on track. When the tool list grows large, tool search loads definitions on demand instead of filling the context with all of them.
Augmented LLM or agent
An augmented LLM is one model call with capabilities attached. It becomes part of a workflow when code chains several calls together, and part of an agent when it runs in a loop and decides its own next step. The same block, with the same tools, can sit inside any of those designs.
That makes the augmented LLM the right place to start testing. If a single call with retrieval and two tools cannot answer a representative question correctly, adding a loop or more agents will not fix it.
Where it goes wrong
Retrieval misses
The right passage is not found, and the model answers from what it has, often confidently.
Overlapping tools
Two tools that do similar things confuse the model about which to call.
Unbounded memory
Memory that stores everything grows noisy and can carry wrong or injected content into later sessions.
How Motiontic's engineers apply it
Motiontic's engineers build and evaluate the single augmented call first: the retrieval, the tool definitions and the memory scope, tested against real questions from the client. Orchestration comes after the block is reliable on its own.
The three augmentations of an LLM
| Augmentation | What it adds | Common implementation | Typical failure |
|---|---|---|---|
| Retrieval | Facts the model was not trained on | Hybrid search over an index, or agentic file search | The relevant passage is not retrieved |
| Tools | The ability to read and change live systems | Function calling, MCP servers | Wrong tool or invented arguments |
| Memory | State that outlives one request | A memory tool, notes files, a database of facts | Stale or wrong memories reused later |
Common questions
What is an augmented LLM?
A language model given retrieval, tools and memory, so it can look things up, act on external systems and keep information between requests. Anthropic describes it as the basic building block of agentic systems.
Is an augmented LLM an agent?
Not on its own. An augmented LLM is a single model call with capabilities. It becomes an agent when it runs in a loop and chooses its own next step based on what its tools return.
Is RAG an augmented LLM?
Retrieval-augmented generation is one augmentation: retrieval. A full augmented LLM in Anthropic's sense also has tools and memory, though many products only need one or two of the three.
Related
AI agent
A model that picks its own next step and tool, in a loop, until the job is done.
Retrieval-augmented generation (RAG)
Fetch the right passages at query time, then answer from them with citations.
Tool calling
How a model asks software to do something: a named tool, typed arguments, a result.
Agent memory
Session state for the task at hand, plus stored knowledge the agent recalls later.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.