Agent Field Guide

Agent Field Guide is a Chinese interactive learning site that takes a reader from the vocabulary of agents to a first verifiable prototype. It is organized as seven chapter views, with a content-operations task running through the examples. The output is a working guide and a single-file offline edition, not a deployed production agent.

Agent is often presented as a label attached to a chatbot, a workflow, or a collection of tools. For a beginner, that blurs the decisions that matter: which steps are fixed, where the model chooses an action, what a tool is allowed to do, and how to tell whether the result is any good. A list of definitions alone does not make those boundaries usable.

The guide makes the tool loop observable before asking the reader to build one. It separates a fixed workflow from an agent's dynamic decisions, then asks the reader to choose the smallest route that fits a real task. The design card records the task, tools, permissions, stopping point, and evidence of success. The aim is to make a modest end-to-end test possible before adding complexity.

The site combines short lessons with five deterministic execution simulations, a seven-step practice checklist, a six-question self-check, and a design card that can be exported as Markdown. Progress stays in the current browser. A downloadable minimal Python example has a dry-run path and a separate real-model path that requires the reader's own configuration. An offline HTML edition includes the code example and front-end assets.

Local interaction checks covered all seven chapter views, concept controls, the five simulation branches and resets, the design card's persistence and escaping, export, the quiz, the theme switch, and the offline page. The Python example passed its dry-run self-check. These checks verify the guide's behavior; they do not measure learning outcomes.

The simulations are scripted, and no real model API call was made during validation. The learning state is browser-local rather than synced between devices. A full visual browser review and a study with learners were not completed, so clarity and instructional effectiveness remain open questions.

Use the guide with one real content task: let a learner complete the design card, build the smallest runnable loop, and compare its output with a simpler baseline. Revise the lessons from the points where that attempt actually breaks down.