Claude Code Lead's Exit Reveals 3 Shifts in AI Coding Workflows

What Cat Wu Actually Said (Fact Recap)

Two years ago , Claude Code was just an internal tool. Cat Wu had it build a reinforcement learning environment, "guiding it for dozens of hours without writing a single line of code herself" — it worked, which drew her to join the team.

At launch (early 2025) , most users manually approved every permission prompt, mainly using Claude for Q&A and writing tests.

Today , the default is to trust Claude's judgment and let it auto-approve . Claude can now handle large-scale code migrations and even build new products.

She thanked the teams behind Claude Code, Cowork, and Tag .

In short: this post is the official footnote to the AI programming "trust curve."

From manual approval to default auto-approval, not a feature update but a workflow paradigm shift

Signal 1: The Trust Curve Matters More Than Model Version Numbers

Notice she describes not "the model got stronger" but whether humans dare to let go . Manual approval → auto-approval: once this switch flips, efficiency jumps by orders of magnitude.

This echoes my previous piece on vertical agents emphasizing "with verification": you only dare auto-approve if you have a way to judge whether it did the job right. Without a verification loop, auto-approval is running naked.

Signal 2: Anthropic's Agent Trio

The three teams Cat Wu thanked are actually the three Agent product lines Anthropic has rolled out — divided by workspace , not by model:

Three agents divided by workspace: repo for Code, files/business for Cowork, Slack for Tag

Claude Code : For engineers, works in repo/terminal/IDE, outputs code diffs, tests, PRs.

Claude Cowork : For knowledge workers, works in files/browser/cloud, outputs reports, spreadsheets, PPTs, organized file sets, includes computer use (research preview).

Claude Tag : For teams, works in Slack channels, @Claude to delegate tasks, runs asynchronously, shares context, can "proactively" follow up — effectively embedding the agent into team collaboration flow.

Signal 3: 65% of Code Now Written by Agent

This is the hardest data point in the piece. Anthropic disclosed when introducing Tag:

Internal version of Claude Tag has already generated 65% of the product team's code.

Don't dismiss it as PR talk. The implication: a top AI company is already using agents to produce the majority of its code, and this "dogfood first" rhythm is often a preview of your workflow six months later. When a company turns agent-written code from "experiment" into "default," the industry inflection point has passed.

Your Decision Hook: How to Choose Among the Three

Look at where your authoritative context lives:

Source of truth is the code repository → Claude Code.

Source of truth is files and business systems (reports, research, docs) → Claude Cowork.

Collaboration and ownership happen in Slack channels , need multi-person visibility and handoff → Claude Tag.

A deeper move: are you still manually approving every AI operation? Cat Wu's retrospective shows that once the trust curve crosses the threshold, efficiency gains are orders of magnitude — but the prerequisite is your verification loop (echoing the "with verification" from the vertical agent piece).

Local-First Counterpoint

All three are cloud-based Claude; data leaves your domain. As I've repeated in previous pieces: you can absolutely run 70B on M5 Max 128GB, 405B on M5 Ultra 512GB locally, building your own "local Code / Cowork / Tag" prototypes — for privacy-sensitive scenarios, local is the answer. The agent paradigm takes shape in the cloud, but the landing point can be your Mac.

Cat Wu didn't say what's next, only replied to someone asking if she'd take six months off: "postagi!" — a joke hinting she'll still work on AGI-related things. The story of the Agent trio is just beginning. Which one do you want to plug into your workflow first?