PROJECT ONE - A MINIMAL HARNESS

EXAMPLES Example 1: The core loop in about fifty lines. A learner who wants to see the idea with nothing else in the way writes a bare loop of roughly fifty lines. It calls the model, checks whether the reply asks for a tool, runs the tool and appends the result if so, and repeats until the model says it is done or a step limit is hit. Nothing in it is clever, which is the point, because this is the smallest thing that deserves the name agent. The main_loop.py file in this chapter is that same skeleton with logging and error handling wrapped around it. Example 2: A first registry with two tools. Before building the lab, a learner might practice in a scratch folder named project_files with a registry holding just read_file and list_directory. Each tool has a simple name, a one-sentence description, and a handler function, and nothing more. Keeping the first registry this small makes it easy to see that a registry is only a dictionary of names pointing to functions. The lab version simply uses three tools instead of two. Example 3: A single hard-coded permission rule. The same learner writes one rule, which is that reading is allowed only inside the project_files folder, and enforces it before every read_file call. Then the learner asks for a path that climbs out with two dots and watches the gate refuse. In the lab the same idea appears as permission_gate.py, which guards the resumes and reports folders instead. Example 4: The first observed run. The learner asks the practice agent what files are in the project. The agent calls the listing tool once, reads two files, and produces a correct summary. It is a small result, but it is a satisfying milestone, because every later chapter is built on this loop working. In the lab your equivalent milestone is the moment five correct reports appear in the reports folder.