Verify agent output before it ships
Run one agent in an isolated worktree and release its staged output only after a second agent verifies the work.
Use this when
Use this when a recurring repository task should run unattended but one agent must not be allowed to generate and approve the same output.
How it runs
- Set the retry limit, wake the due Loop Harness task, and create an isolated worktree from the approved source revision.
- Have the primary Claude session stage one bounded result without publishing it.
- Have a second Claude session inspect the staged work against explicit acceptance criteria.
- Ship on a pass; otherwise preserve the findings, publish nothing, and retry only until the preset limit.
Done when
✓ Only independently verified output ships. A second-agent pass releases the configured output; a failed verification preserves evidence and produces no external change.
Why it works
Workspace isolation limits interference, and the second-agent gate separates generation from approval. The result can run repeatedly without relying on one session's confidence.
Implementation note
The source implementation uses Loop Harness, git worktrees, and separate model sessions. Start with read-only tasks, test one run first, cap runtime and retries, and grant only the tools each agent needs.
Source: Forward Future ↗graded C · 60/100 — how grades work →
More git loops
Ship a PR until green
Implement a change, open the PR with gh, then keep fixing CI failures until every check passes, all in one goal loop.
Commit message hygiene on a branch
Rewrite the commit messages on your feature branch to conventional-commit format with meaningful bodies before opening the PR, leaving the code untouched.