A planner, coder and reviewer team passed 25.7% of hidden tests; one agent passed 56.8%

The author ran Codex CLI 0.157 with the hosted gpt-5.6-luna model at medium effort on six feature tasks in shopizer, an open-source Java shop: 144 runs graded by hidden acceptance tests. One agent passed 56.8%. The planner, coder and reviewer team passed 25.7% at $0.107 per run against $0.073. The author's explanation is the hand-off: only the planner saw the task, and the coder's code compiled but broke the stated interface, such as a plain array where the task asked for {"count": ..., "items": [...]}. When every agent was also given the task file, the team reached 54.1%, which the author reports as level with one agent (-2.7 points, interval -13.7 to +8.2) at about 1.6 times the cost.

Every one of the 24 final reviews that approved a change was followed by hidden-test failures, and no agent wrote a test in any of the 144 runs. In the author's words, "the agent reviewer approved broken code every time it approved anything." The limits are stated: one model, one effort level, six features in one codebase, and team instructions the author wrote. Codex stores inter-agent messages encrypted, so the hand-off loss is inferred from the task-file follow-up, which was designed after the main results.

Why it matters: If a team of agents is on your roadmap, the requirements have to sit where every agent reads them, not only in the planner's prompt. An agent reviewer's approval is not a gate on this evidence, and a single agent at the same budget is the baseline to beat. This is one measurement on one setup, so run the comparison on your own tasks before you rely on it.