Language models are, paradoxically, bad at things a calculator does effortlessly: exact arithmetic or factual lookup (Alammar & Grootendorst, 2024, p. 218; Schick et al., 2023). The obvious fix is to give them tools. The mechanism is simple: the model only produces text, so it is trained or instructed in the system prompt that, when it needs a tool, it should write a call in a fixed format (for instance, a function name and its arguments). The program wrapped around it detects the call, executes it and pastes the result into the context; the model keeps writing from there.
Calculator(400 / 1400) → 0.29
Toolformer showed that a model can learn on its own when to call which tool (a calculator, a question-answering system, a search engine, a translator, a calendar), which arguments to pass and how to use the response, in a self-supervised way and with just a handful of demonstrations per tool (Schick et al., 2023). Chaining many calls with intermediate reasoning is what defines an agent.
Millière & Buckner (2024b) note that this builds hybrid neurosymbolic systems, in which language or code serves as a universal interface between connectionist and symbolic components; with a code interpreter or Wolfram Alpha, for example, performance on hard math improves considerably (2024b, sec. 3.1). Mahowald et al. (2024) read it in the light of neuroscience: language and formal reasoning are distinct capacities, and it makes sense for them to rely on separate mechanisms. It is a point in the debate that Santelli (2026) takes up: for some critics, an LLM's need for tools proves that growing networks is not enough; for others, it is no more surprising than a person doing better with an abacus, since the "processor" is still the grown network. The underlying question is where to draw the boundary of the cognitive system: is the calculator part of the agent's mind or an external instrument?