Tools in agentic AI and how agents pick them

Tools are typed functions the agent can invoke: search_web, query_database, send_email, run_code, etc., each with a JSON schema describing inputs/outputs. The agent's policy decides which tool to call: by default, the LLM itself emits a tool_call field tuned during fine-tuning or via prompt scaffolding (system prompt listing tools and rules). Better agents also use: (1) explicit selector prompts ("given the user's request and the available tools, output a JSON plan"); (2) search-style planners that retrieve tools by semantic similarity to the user's need; (3) constrained decoding (JSON-schema guided generation) to guarantee well-formed calls. Selection quality depends on (a) tool description clarity, (b) tool routing logic, (c) error feedback. Senior nuance: parallel tool calls (multiple tools in one step) are common for "gather information" substeps, but require careful deduplication.