AI-agent projects lack the deeper testing coverage their risks require

A review of 157 open-source AI-agent projects found ordinary tests in 87.3% of repositories, but much thinner end-to-end, adversarial, multi-route, and prompt-injection coverage. Benchmark evidence that multi-step reliability can collapse despite strong single-turn performance reinforces testing real actions, tool boundaries, permissions, downstream effects, and end states rather than relying on fluent responses.