Textbook of AI

Tool use is the foundational concept underpinning every modern AI agent. A pure LLM is a stateless function from text to text; tool use extends its action space to include arbitrary external programs.

Why tools

LLMs have three irreducible weaknesses that tools repair:

1. Stale knowledge, training data has a cutoff; tools (web search, databases) supply fresh facts.
2. Brittle arithmetic, transformers struggle with multi-digit multiplication; a calculator tool is exact.
3. No side-effects, language models cannot send an email or commit code; tools provide actuators.

Categories of tool

Mechanism

Three generations of tool-use plumbing have been deployed:

1. Prompt-only (2022), the model emits a tool call as plain text, e.g. Action: Search[query], parsed by regex (the original ReAct format).
2. Native function calling (2023+), provider APIs accept a tools list and emit structured JSON (function calling).
3. Model Context Protocol (MCP) (2024+), Anthropic's open standard letting any tool server be discovered and invoked uniformly across clients.

Relationship to other patterns

• ReAct is the loop that uses tools.
• Function calling is the wire format for tool invocation.
• Toolformer is a training method that teaches an LLM when to use tools.
• Agentic RAG is tool use applied specifically to retrieval.

Frontier capability

By 2025 the SOTA models on agent benchmarks (SWE-bench, GAIA, $\tau$-bench) all rely on multi-step tool use. The bottleneck has shifted from "can the model call tools" to "can the model orchestrate fifty tool calls without losing the thread", a question of context management, not interface design.