Zylos research says chat UIs can't handle multi-step AI agents

Zylos's research report argues that the plain chat interface, fine for turn-by-turn conversation, breaks down once agents run multi-step, autonomous workflows: users can't see what happened, can't intervene mid-task, and can't recover cleanly from errors. It cites Augment Code's finding that agent sessions without live progress panels saw triple the abandonment rate of sessions with them, even when output quality matched, and an Anthropic study showing experienced Claude Code users auto-approve over 40% of actions, more than double the rate for newcomers. The piece maps two emerging protocols, CopilotKit's AG-UI for streaming agent events to the frontend and Google's A2UI for declaring UI components in JSON, and lays out three generative-UI patterns trading agent autonomy against developer control. Its concrete recommendation is a separate, auditable activity panel apart from the conversation thread, plus per-user approval settings that loosen as trust builds.