What changes when the whole team can steer the coding agent?

In Simon’s conversation with Cat Wu and Thariq Shihipar, the striking number was 65%: the share of the Claude Code product engineering team’s pull requests landed by its internal Claude Tag. That figure was specifically about this team, not all of Anthropic.

An agent in a shared Slack channel can take a first pass, show the result, receive design feedback and then hand the work to engineering. Colleagues also see how one another use it. Effective practices spread through ordinary work.

Review remains uneven by design. Critical areas have human code owners. Some outer layers rely more heavily on automated review, after months of iteration. When an incident escapes, the responsible change becomes material for a regression evaluation.

Features must also demonstrate retention among internal users before wider release. Cheap implementation has not eliminated the question of whether anyone wants the result.

This is a more useful picture of adoption than counting generated lines: shared context, visible decisions and a feedback loop that improves the system doing the work.