Building with AI · Agent UX patterns
You can explain the difference between per-action and plan-level approval, and why the field moved away from the former.
The first agent products that shipped tool use — Claude with function calling, the earliest Cursor agent mode, ChatGPT plugins — asked the user before every tool call. Read this file? Approve. Edit it? Approve. Run the tests? Approve. The pattern made sense from a safety standpoint: humans-in-the-loop on every action.
It didn't survive contact with real users. Anthropic's Trustworthy Agents in Practice reports developers approve 93% of permission prompts in Claude Code's auto mode. At that approval rate, the prompt is not a control surface — it's a speed bump that trains the user to click "approve" reflexively. The 7% rejections are then drowned in the muscle memory.
Plan mode is the replacement. The agent emits its full plan up front. The user reads it once, edits it, approves it, and then the agent executes the whole thing without further prompts. The human attention is concentrated at the moment when it can actually do work — reading the plan — instead of spread thin across 47 micro-approvals where it can't.
This chapter teaches the pattern as it ships in real products today, and the design judgment calls that come with it: when per-action approval still makes sense (rare, but real), and how to design the approval prompt itself so "approve" or "reject" is a quick, informed decision.
Type: multi-choice
Chapter contains 3 lessons.