Agent-written code is not the problem. Unreviewed agent loops are.
The Claude Code team described how they review code when agents produce most of it: review starts before the PR, Claude must run tests, builds or screenshots, another subagent reviews the diff, and rollback is treated as part of the system. Their reported result is interesting: from June to September, merged PRs more than doubled while the share reverted fell by more than 40%.
My angle for small business automation
This is exactly how I prefer to build AI pipelines around invoices, leads, warehouse updates or support replies. Do not let an agent just "change the script" or edit an n8n flow and hope. Give it a verifier: sample invoices, expected JSON, a dry-run mode, a database diff, a second Claude review pass, and a fast rollback. Boring checks pay for themselves because they protect the hours automation gives back.
I'd say this honestly: this is not a license to auto-merge everything. For payments, legal summaries, client emails and stock movements, humans still own the final business decision. But the routine reading - syntax, edge cases, regression checks - should be done by machines first, not demo videos.
Why it matters: AI agents become useful in business only when they are wrapped in verification. The win is not "Claude wrote code" - the win is fewer manual checks without more incidents.
Source: Hacker News