Build AI Agents in Python: A Practical Travel Agent

AI agents become useful when model judgment works inside application-controlled boundaries. A prompt alone does not decide which inputs are valid, which tools are authorized, which evidence may be trusted, what output shape downstream software can accept, or whether a system may take an external action.

This practical course teaches those engineering decisions by building IST Travels, a planning-only travel agent in Python. You begin by separating chatbots, fixed workflows, and agents. Then you build one focused planner with typed requests, bounded destination evidence, structured plans, and an optional OpenAI Agents SDK path.

Each module advances the same working system. You add flight, lodging, activity, and weather tools; reconcile compatible evidence; resolve current preferences against approved memory; introduce specialist agents only when responsibilities genuinely split; and preserve one accountable manager for the final answer.

The course deliberately breaks the system as it grows. You will diagnose unsupported destinations, incompatible evidence, memory-precedence regressions, unsafe reservation intent, sensitive input, shared-budget failures, evaluation regressions, provider timeouts, invalid API requests, and unsafe production configuration. Every failure has a named owner and a reproducible checkpoint.

The final capstone exposes the complete workflow through FastAPI with health, readiness, request identity, typed responses, explicit limitations, and real local HTTP smoke evidence. IST Travels never reserves, purchases, holds inventory, or claims that sample prices are live.

This is a code-first build-along with real files, commands, tests, traces, failures, diagnosis, and cumulative verification. The core path runs locally with deterministic sample providers, so learners can understand the architecture before choosing whether to configure a paid model or external provider.