CODING AGENT HARNESSES

EXAMPLES Example 1: A terminal-based versus editor-based coding agent. A developer asks a terminal-based agent to find why a program crashes. The agent searches the files, reads the relevant code, runs the program in the shell, and reads the error message. Another developer uses an editor-based agent that suggests a fix inline, showing the changed lines highlighted so that she can accept or reject them. The first approach has greater reach and needs stricter permissions, while the second keeps the person closer to each change. Example 2: A reliability constraint. A team requires that the agent first write a failing test that reproduces a reported bug. The agent writes a test that fails, which confirms the bug is real. Then it changes the code and runs the test again, and it passes. The team has objective proof that the fix works, instead of trusting the agent's claim. Example 3: A verification gate. An agent finishes a change and reports that everything works. The pipeline automatically runs the existing test suite, and two older tests fail. The change is rejected and sent back with the failure messages. The agent discovers that its change broke an assumption elsewhere, repairs it, and the gate passes on the second attempt. Example 4: A lesson from agent-built codebases. One team gives an agent a vague assignment to modernize the whole billing system, and the result is a large, tangled change that nobody can review. The team then splits the work into twenty narrow tasks, each with a clear goal and a test that must pass. The new approach produces smaller changes that are reviewed quickly, and far fewer problems reach users.