Robots Atlas>ROBOTS ATLAS
Agents

Tool Use

2023ActivePublished: 29 September 2026Updated: 29 September 2026Published
Key innovation
A language model’s ability to call external tools (APIs, search, a code interpreter) during generation to go beyond its parametric knowledge.
Category
Agents
Abstraction level
Pattern
Operation level
Agent runtimeInferenceTooling
Use cases
Web search and RAGCode execution / computationCalling APIs and integrationsWorkflow automation (agents)Controlling robots via tools

How it works

The model is given descriptions of available tools (function schemas) and, during generation, can emit a call with arguments. The agent runtime executes the tool, returns the result as an observation, and the model continues — often in a reason→act→observe loop (ReAct). Modern models are trained for function calling, and protocols like MCP standardise how tools are connected.

Problem solved

An LLM alone has frozen knowledge, cannot compute reliably and has no access to the world. Tool use gives it fresh data, exact computation and the ability to act.

Components

Tool schemaInterface

Function description: name, parameters, types — the model knows what and how to call.

Function callingUse decision

Structured model output naming a tool and arguments.

Execution runtime and observationFeedback

The environment executes the tool and returns the result to the context.

Evolution

Original paper · 2023 · Timo Schick
Toolformer: Language Models Can Teach Themselves to Use Tools
Timo Schick, et al.
2022
ReAct interleaves reasoning and acting with tools
Inflection point
2023
Toolformer: a model learns by itself when to call APIs
2024
Standardised tool connectivity (MCP) and native function calling