Beyond "Prompt and Pray"

A language model predicts tokens. An agent acts — and has to answer for what it did.

That gap is where most agentic deployments fail. The prototype works, the demo goes well, and then the system reaches production and nobody can say why it took a particular action, what evidence it relied on, or whether it would do the same thing again. The usual response is a longer prompt and a hopeful deploy.

This book takes the other route. It treats an agent as a system you architect rather than a model you coax: bounded actions, traceable evidence, and failure that surfaces early instead of arriving as a customer complaint.

You build it up in layers — perception, planning, tools, memory, evaluation, runtime governance — with Python examples throughout. The capstone is a complete banking complaint agent, the kind of workload where a wrong answer has a regulator attached to it.

Written for engineers and architects shipping agents into environments where “it usually works” isn’t a standard anyone accepts.