CODING AGENT HARNESSES

CHAPTER SUMMARY A coding agent reads, writes, and changes programs, and it is a good teaching case because code can be checked objectively. Terminal-based, editor-based, and cloud-based agents trade reach against closeness of review, and all need the same harness duties. Constraints such as narrow scope, small steps, protected files, diff review, and test-first work make generated code more reliable. Verification gates automatically reject work that fails tests, linting, type checks, builds, or security scans, regardless of the agent's confidence. Teams report that clear, narrow, verifiable tasks work far better than vague, large ones, and humans remain essential for judgment. KEY TERMS REVIEW Coding agent: an agent whose main job is reading and writing programs. Terminal: a window for typing text commands to the computer. Shell: the program that interprets typed commands. Diff: the list of exact lines added and removed by a change. Test: a small program that checks whether code behaves correctly. Test-driven development: writing a failing test before writing the fix. Linting: automatic scanning of code for style problems and likely mistakes. Type checking: confirming values are used in the ways the code declares. Verification gate: an automatic checkpoint that sends failing work back.