Claude Code agent loops need a design review

What it does

Looper acts as a pre-flight design layer for Claude Code, invoked with /looper. It tightens fuzzy automation ideas against built-in best-practice rubrics and scaffolds a portable loop spec — complete with typed verification gates and a cross-model reviewer — before any runner starts changing files. The output is a runnable handoff prompt for the current session or a thin run-loop.py you own and edit.

The interesting bit

Instead of replacing Claude Code’s native /goal or /loop, Looper sits upstream as a deliberate design coach, even recommending a different model family sit in the review seat so the worker isn’t grading its own homework.

Key highlights

• Forces checkable verification classified as programmatic, judge, or human before execution.
• Generates a portable, versionable spec (loop.yaml, loop.resolved.json) plus a thin run-loop.py you can edit.
• Includes termination guards: iteration caps, revision limits, no-progress detection, and budget advisories.
• Recommends a cross-model reviewer or judge to avoid self-evaluation blind spots.
• Runs immediately in-session via a handoff prompt or later via the optional Python runner.

Caveats

• Budget caps are advisory limits visible to the operator, not hard enforcement unless you add accounting around the configured CLIs.
• It explicitly does not handle durable orchestration, cron scheduling, or retries across process restarts — design only.

Verdict

Developers who repeatedly watch Claude Code spiral on poorly defined tasks should install this immediately. If you already know exactly what your loop needs and just want a cron job or session persistence, stick with Claude’s native /goal and /loop instead.