Chapter 1: Why Trust?

This guide is about building AI agents in Python (on Azure, with GitHub Copilot) that you can actually trust in front of customers you care about.

There is a tension at the heart of it: the difference between trusting a local hawker uncle who has cooked the same dish for thirty years, and trusting a chatbot you met five seconds ago. Closing that gap is the whole job.

Why trust is the hard part

A polished demo is easy; that is what demos are for. The hard part is the gap between a demo and a system that strangers depend on every day. An agent that dazzles on a laptop can still hallucinate facts, leak data, fail to explain itself, or fall over under real traffic. "Trustworthy" is not a vibe. It is a set of concrete engineering properties you can design for, test, and observe.

This guide gives you two things:

• Developers get a copyable pattern (small agents, guardrails, observability) and three working examples to run.
• Decision makers get a vocabulary (transparency, safety, reliability, observability) and a sense of what each one costs and buys.

The promise

No magic mega-agent. Instead: small, specialised, well-governed agents that collaborate. That is the pattern that actually ships.

Key terms

• Agent: an LLM-driven program that can call tools, make decisions, and produce an answer, often over multiple steps.
• Trustworthy: explainable, safe, reliable, and observable. The rest of these chapters define each property concretely.

Do this next

1. Clone the repo: README.md has setup steps.
2. Skim the source layout under src/merlions/: three agents, their tools, their policies, and an eval suite.
3. Read the chapters in order, or jump to the agent that matches your use case.

📺 Go deeper: the Microsoft Build 2026 sessions that inform this guide are listed in talk/context.md. Start with BRK250 (observe and control agents across any framework).