Severe: Claude Code runs a multi-hour multi-agent pipeline on a UI-only change and ignores explicit ordering instructions

Summary

Claude Code turned a UI-only change into a multi-hour, multi-agent pipeline. The change had no database migration, no new page and no new endpoint. The user had already complained about slowness earlier in the same session, and Claude Code did it anyway. This is a serious and recurring failure to scale process to risk. It burned hours of the user's time and a large amount of tokens on work that should have taken a fraction of that.

Environment

• Claude Code 2.1.296, macOS
• Model: claude-opus-5-5
• Private Nuxt web app, admin panel

The task

The task was pure layout and UI:

• Move phone-only page actions into the mobile top bar (search, filter, action keys, create).
• Re-lay out two pages for phones.
• Widen one settings page on desktop.

What Claude Code did

• It split the task into two PRs and ran the full heavyweight pipeline on each: a background implementer subagent, then a separate reviewer subagent, then a fix round, with a CI wait between every step. It then queued an extra cross-model overall review.
• Agent time alone:

| PR | Step | Time |
|---|---|---|
| PR 1 | implement | ~50 min |
| PR 1 | review | ~15 min |
| PR 1 | fix | ~26 min |
| PR 2 | implement | ~31 min |
| PR 2 | review | ~12 min |

A second fix round on PR 2 was underway and was stopped only after the user objected. That is roughly 2.5 hours of agent time before CI waits, for a layout change.

• The implementers wrote failing-first tests for layout tweaks and ran 100+ scripted browser checks per PR.
• Earlier in the same session, the user explicitly said: write the design rule into the design doc first, then start designing. Claude Code published design candidates and started exploring code first. It also held a long queued request behind a small visual pick.
• The user complained about slowness earlier in the session ("too slow"). Claude Code saved a memory note about it, then repeated the same pattern on the very next task.

Expected

• Scale process to risk. A UI-only change with no schema, routes or endpoints should be one PR and one implementation pass. Review should run in parallel with CI. Low-severity findings should go into the PR description instead of costing extra fix rounds. Done in well under an hour.
• Follow explicit ordering instructions literally. "Do X first" means X is the first action.
• Change behaviour immediately after a complaint about speed, for the rest of the session, not just write a note and carry on.

Impact

• Hours of wall-clock time lost on trivial UI work.
• The user had to step in repeatedly.
• Significant token cost.

This makes Claude Code hard to trust for day-to-day product work. Please treat it as a high-priority behaviour problem, not a one-off.