Observed facts: The author ran the same payment-incident task against two copies of one local repository, using Koda Pro based on GLM 5.3. The solo run scored 7/8 on an external hidden HTTP evaluator; the lead-agent-plus-four-subagents run scored 8/8. Both runs passed their own tests and builds. The missed scenario was a new checkout for an already-paid order: the solo version returned 200, while the team version returned 409.
The mechanism: The four roles investigated distinct uncertainty areas—frontend retry behavior, payment backend and database transactions, webhook processing, and adversarial QA—and were read-only. The lead agent alone integrated changes. This creates parallel inspection without parallel edits; the external evaluator then tests business behavior beyond the agents' self-reported success.
Why now: Coding agents are increasingly presented as teams, so the operational question is not simply whether delegation produces more code, but whether it improves discovery and verification on a bounded task.
First implication: For teams testing agent workflows, assign parallel agents to independent risk surfaces and preserve an independent evaluator for business invariants.
The evidence boundary: This is one author's comparison on one repository and task, with one model configuration and two runs. The different test counts and implementation sizes also make the results unsuitable as a controlled estimate of general multi-agent advantage.
First watch condition: A follow-up comparison across multiple repositories and repeated runs would show whether the hidden-test gap persists when task and project variation are introduced.
The broader takeaway is about verification bandwidth, not a blanket recommendation to add agents. Here, an adversarial role anticipated a paid-order edge case, and the evaluator made the difference visible. The new evidence is a concrete hidden-test miss and catch, not a claim that the same arrangement will improve every coding task.