Claude Code shows that agentic coding is also an organizational change

Simon Willison published an edited transcript of a fireside chat with Cat Wu and Thariq Shihipar from the Claude Code team. The discussion covers Claude Tag, Fable, coding agent security, evals, tool design and Anthropic’s internal use of its own tools.

Claude Tag has moved into the review queue, not just the editor

The strongest detail is operational: Claude Tag, the collaborative Slack integration, now lands 65 % of product engineering PRs for the Claude Code team, according to the team. That moves the coding agent from a local helper into a shared workflow.

Willison also highlights that critical changes to Claude Code are still reviewed manually, while the team increasingly uses automated code review for the outer layers of the product. Anthropic ships features to its own employees first and only continues with features that show retention in that cohort.

The hard skill is deciding what work to give away

The conversation shows a shift in the developer role. If the agent can handle routine implementation, the job is no longer just to ask for a small diff. The team has to describe product intent, business context and boundaries faster.

That matters for engineering leads. Review moves from watching every permission prompt to choosing the right tasks, inspecting risky changes and measuring whether the output survives real use.

A shorter prompt does not remove the need for judgment

The team also said that adding examples to a system prompt is no longer best practice for models such as Fable 5 or Opus 4.8. The Claude Code system prompt was recently reduced by 80 %.

That sounds simpler, but for teams it mostly moves complexity elsewhere. Fewer prompt rules put more weight on evals, safety boundaries, tool use auditing and the ability to notice when an agent is confidently moving in the wrong direction.

Adoption will show up in review quality, not launch counts

The next signal is the quality of the review queue. If Claude Tag creates most PRs, it will become obvious whether humans are reviewing architecture and product impact or merely approving clean looking diffs.

The other question is whether Anthropic can translate its internal culture of working in public Slack channels into ordinary companies. Without that culture, the agent becomes another noisy bot in the notification pile.

Lilith's verdict

Claude Code is not acting like a faster keyboard here. It is the colleague already standing at the sprint board and taking tickets. The question is who can still tell it no when the diff looks clean.