← Field Notes

Sighting · August 26, 2026

by Daniel Clark

Daniel Clark writes about the interaction patterns and behaviors emerging between humans and agents. His beat is how design, product and engineering keep ahead of the curve to make the work between agents and humans safe, secure and governed.

Claude Code makes you approve the plan, not each command

Claude Code's plan approval moves the decision from every action to the shape of the work, and that changes what approval means.

Approval is usually asked for one action at a time: the agent wants to run a command, you say yes, it wants to edit a file, you say yes again. Claude Code ships that too. But the interesting decision is one level up: Plan Mode, where nothing executes until you've approved the plan as a unit.

This changes what you're consenting to. Per-action approval asks "is this command safe?" — a question you'll answer reflexively by the fortieth prompt. Plan approval asks "is this the right shape of work?" — a question worth a human's attention exactly once per task.

The scope of a yes is the other quiet innovation. "Allow once / allow always / per-session rules" makes authorization a first-class, editable thing rather than a moment that evaporates. You can inspect what you've granted. Compare Codex, which reaches the same goals through modes — read-only, auto, full access — set globally rather than accumulated per rule.

Two vendors, same problem, opposite grain: Claude Code scopes consent per plan and per rule; Codex scopes it per mode and per command. The pattern page tracks both, part by part.