Hands-On AI Agent Engineering

Welcome to the ConfPlanner workshop — a hands-on course in engineering AI agents whose behaviour you can read, bound, and therefore trust in production. Trust here is a concrete, narrow thing: you can inspect what the agent planned, and there are hard invariants it cannot violate — not a claim that its model's judgement is always right (evaluating output quality is a separate discipline this workshop deliberately scopes out). ConfPlanner is a conference-planning agent built around a Goal-Oriented Action Planning (GOAP) core, and the workshop walks you through building, testing, and operating it. The central idea is the dual harness: a build-time harness that constrains how the agent's code is written and changed, and a run-time harness that constrains how the agent behaves once it is running — guardrails, invariants, and observability that keep the agent inside the lines whether a human or another agent is driving it.

How to use these docs

These docs follow the Diataxis framework, which splits documentation into four quadrants by the reader's need. Pick the section that matches what you are trying to do right now.

• Tutorials — learning-oriented. Step-by-step lessons that take you from zero to a working ConfPlanner. Start here if you are new.
• How-to guides — task-oriented. Recipes for getting a specific job done: running with a real model, enabling tracing, adding an action or guardrail, driving a change through the harness.
• Reference — information-oriented. Precise technical descriptions of the domain model, configuration, catalog schema, and CLI. Consult it when you need exact details.
• Explanation — understanding-oriented. The why behind the design: DICE, GOAP, guardrails and invariants, the dual harness, the eight habits, and right-sizing models.

Source code

The full source for ConfPlanner and this workshop lives on GitHub: Habitat-Thinking/hands-on-agent-development.