Part 1 of 4 · 1 min read
Overview
Agents run longer and act more than chat models, so the interface must carry trust, control and recovery. Interviewers ask "how would you show an agent's progress without overwhelming the user?"; the answer is layers, attention routing and control, backed by an event stream (Level 11 module 4).
About the example, and what is real
Real: the status-line, attention, change-list and grouping logic, the approval tiering, batching and expiry logic, the undo log and delayed-send queue, the confidence-wording and ask-or-act rule, and the 19 tests in code/mini-ui. Simulated and assumed: the event streams (random tool calls, a 5 percent error chance), the attention-fatigue model (95 percent initial attention, 7 percent decline per prompt) and the ambiguity model (three readings with noisy scores, true reading centred at 0.55 against 0.30 and 0.15, noise 0.15). Not real: no browser UI, no users; nothing here measures how real people behave.