Textbook of AI

Abstract. Anthropic's 2025 engineering report on production patterns for LLM agents. Surveys the architectural vocabulary that emerged from real deployments, workflows versus agents, augmented LLMs, prompt chaining, routing, parallelisation, orchestrator-worker patterns and evaluator-optimiser loops, and argues that most production "agents" are best implemented as fixed workflows rather than open-ended autonomous loops. The report also documents the security stance the field has converged on for tool-using agents: assume tool output is hostile, log every action at a rollback-compatible granularity, require user confirmation for irreversible actions, and bound authority via capability tokens at the system layer rather than via the model's own judgement.