Mastering AI Engineering

Mastering AI Engineering treats an AI agent as what it really is: a software system with one unpredictable component inside it, not a prompt that happens to work in a demo. It draws one hard line through every chapter: the harness is the runtime that owns the run, the loop is the control mechanism that decides the next step, and the loop proposes while the harness disposes. The book follows one reference agent, Tally, an accounts payable agent at a fictional distributor, as it grows one capability per chapter, each one arriving because something broke and every failure reproducible from the companion code. It covers harness engineering, run envelopes, budgets, checkpoints and cancellation, loop engineering with explicit state and progress detection, context engineering with token budgets, provenance and freshness, tool design with contracts, error taxonomies and idempotency keys, memory architecture kept subordinate to the system of record, orchestration patterns and when not to use multiple agents, guardrails and permissions mapped to the OWASP agentic risks, human approvals bound to the evidence they were given on, evals with confidence intervals and release gates, and observability with OpenTelemetry through production operations and incident response. More than thirty architecture diagrams, working Python code with tests, failure labs, production checklists and decision records make every idea something a reader can build and verify.