Unit 17: Coding Agents & Repository Control

A chat response can suggest an edit. A coding agent may be able to inspect files, search a repository, edit multiple files, run commands, and react to tool results.

The difference is not that the agent suddenly became the project owner. The difference is that model output can now cause actions.

Four boundaries matter

Keep these separate:

MODEL CAPABILITY   what the model can propose
TOOL CAPABILITY    what connected tools can do
PERMISSION         what this task allows
AUTHORITY          who decides whether the result enters the project

A tool may technically allow repository-wide edits while the task grants permission to change two files. Tool capability does not expand scope.

You will write an agent task specification, require a plan before action when possible, establish a known repository checkpoint, inspect the resulting status and diff, run tests yourself, classify unexpected changes, and decide whether to keep or reject the work.

No repository-writing agent is required. The fallback path uses a supplied plan and patch so you can practice permission, review, testing, and recovery without granting live write access.

Evidence of control

Your Unit artifact should show one bounded task from specification through review decision. The important proof is not that the agent said “done.” It is the repository evidence supporting your decision after the claim.