An augmented LLM is a large language model equipped with augmentations such as tools. Simon Willison reports the term from an Anthropic article about building agentic systems and singles it out as one he really likes, on the grounds that he has seen people use “agents” for just this — a model with tools attached — which never felt right to him. The term’s value on his account is therefore boundary-drawing: it gives the simplest case its own name, so that “agent” does not have to cover it.
Usage
What Willison praises the article for is defining its terms at all: his standing complaint about “agents” is that the word has many potential definitions while most people using it assume everyone shares the one they picked, and he credits the article with bucking that trend from the start. As he reports it, the article treats “agentic systems” as a parent term and then distinguishes “workflows” — systems where multiple LLMs are orchestrated together using pre-defined patterns — from “agents”, where the LLMs dynamically direct their own processes and tool usage. The augmented LLM is a further term he singles out from the same article, and it names a different kind of thing from either: a single LLM with augmentations such as tools, rather than several LLMs arranged into a system.
The distinction matters most where the word “agent” is being claimed. Willison’s account of the article’s own advice points the same way — that when building applications with LLMs the recommendation is to find the simplest solution possible and only increase complexity when needed, which might mean not building agentic systems at all, and to avoid investing in complex agent frameworks before exhausting direct API access and simple code. Naming the simplest configuration is what makes it possible to say that a given system is only that.
Nothing here is Willison’s own coinage: he is reporting and endorsing someone else’s term, in a link post about their article. The definition available through that source is correspondingly brief — an LLM with augmentations such as tools. The article he is reporting is held directly by this page as well, and is read below.
The Primary Account
The article Willison reports is Building effective agents, which this page now also holds directly, so the term can be read in the words that introduced it rather than only at one remove.
There, the augmented LLM is presented as the basic building block of agentic systems: a model enhanced with augmentations such as retrieval, tools and memory. What Anthropic emphasises about the current generation of models is that they can use those capabilities actively rather than being driven through them — generating their own search queries, selecting appropriate tools, and determining what information to retain. The post’s two implementation recommendations are to tailor these capabilities to the specific use case, and to give the model an easy, well-documented interface to them; Model Context Protocol is named as one way to supply that interface.
Its place in the post’s structure is what makes it a building block rather than one option among several. Everything the post goes on to describe — prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer, and the autonomous agent — is built on the assumption, stated explicitly, that every LLM call from that point onwards has these augmented capabilities available. So the augmented LLM is not the simplest item in a list of architectures; it is the unit the rest of the list is composed from.
That also sharpens the boundary Willison values. Anthropic groups workflows and agents together as agentic systems and separates them by who directs the process — predefined code paths in a workflow, the model itself in an agent (see AI Agent). Reading the augmented LLM as sitting below that split — a single call with capabilities attached, not yet a system in which anything is being orchestrated — is this page’s own inference from its position as the building block; the post places it before the patterns without placing it against the split itself.
Related Terms
- Function calling — the mechanism the augmentation usually consists of
- AI Coding Agent
- Sub-agent architecture