• Yao et al. (ICLR 2023): ReAct
• OpenAI: function calling (Chat Completions examples)
• Anthropic: client tool use and tool results
• JSON Schema: object properties, required and additionalProperties

This is a simulation. Model decisions follow a fixed script, a stand-in for an LLM. The tools run real local arithmetic over an in-memory example workspace; no external service or real file is accessed. Token counts are fixed estimates per message and tool schema, not tokenizer measurements or prices. Schemas and transcripts use the OpenAI Chat Completions format, with one tool call per model turn. The validator implements only the schema keywords used here. Independent lesson, not affiliated with OpenAI or Anthropic.