Instruction following is the capability of a language model to accurately execute the directives given in user or system prompts — including output format, length constraints, topical scope, tone, and behavioral rules. It is a foundational skill for agentic use cases because agents depend on the model reliably obeying structured instructions such as tool-call schemas, output parsers, and multi-step plans. Models are specifically trained and evaluated on instruction-following benchmarks to improve reliability in these scenarios.