For two years, the debate over "will AI replace programmers" evolved from a serious industry anxiety into a stale internet meme. Every few months a new tool emerges to prove "programmers haven't been replaced yet." Copilot reduced your keystrokes, Cursor reduced your documentation lookups, Devin reduced your deployment worries. Each time the conclusion seemed to be the same: AI coding efficiency is rising, but human engineers still sit in the driver's seat.
But look at what happened with Fable 5 this week, and the frame of the question has changed. AI did not replace programmers — the programmer's job description got rewritten. From "person who writes code" to "person who accepts AI's work." From building every feature yourself to @mentioning Claude in Slack and waiting for the PR to arrive.
This is not an efficiency story. It is an organizational-structure story.
Three Leaps: From Autocomplete to Delegating
Boris Cherny, creator of Claude Code, recently mapped AI coding's evolution into three clear stages in an interview. Leap one: AI completes the next line. The Copilot era from 2021 — AI as a "supercharged autocomplete," guessing the second half of what you started. Decision power stays with the human: you accept or reject each line. Leap two: AI writes complete features. 2024-2025 brought agent-style tools like Claude Code and Cursor. You stop completing code and start "delegating": write a function, build an API, refactor a module. AI reads the codebase, runs tests, fixes bugs. One person manages 10 concurrent Claude sessions, building multiple features in parallel. Leap three: the AI team goes to work, the human becomes the acceptor. This is the Fable 5 plus Claude Tag stage. On June 23, 2026, Anthropic put Claude Tag into Slack. You no longer open any IDE — @mention Claude in a channel, describe what you need, and it decomposes the task, calls tools, writes code, and submits a PR.
Boris himself runs dozens of Claude Tag sessions — some running for days, some for weeks. His daily work has become: watching PRs arrive one by one, watching data reports come in one by one. And he dropped a number worth noting: the product team's 65% of code is now written by the internal Claude Tag deployment, and the ratio keeps climbing. Sixty-five percent. That number's meaning goes far beyond "AI's share of code written." Its subtext: the engineer's core output is shifting from "code I wrote" to "AI output I have reviewed and accepted."
Anthropic's Product Matrix: a Three-Layer Architecture
The Claude Tag launch is not isolated — it slots into a deliberate product architecture. Claude Code handles deep coding: the IDE-resident agent that reads repositories, runs tests, and manages complex engineering tasks. Claude Tag handles delegation: the Slack-native entry point where any team member can trigger AI work without leaving their communication tool. Artifacts handles delivery: turning AI output into shareable, editable pages. Cowork handles long-horizon work: cross-application, multi-hour autonomous execution for non-developers. Each layer targets a different depth of AI involvement — and each layer's users are further from writing code than the last.
The progression matters because it reveals Anthropic's thesis: the future of AI coding is not a better IDE. It is AI embedded in the tools people already use — communication tools, project-management tools, document tools — so that "using AI" stops being a separate activity and becomes an invisible layer in existing workflows. Claude Tag in Slack is the most aggressive version of this thesis: it removes the IDE from the loop entirely. A product manager who has never opened a terminal can now trigger a code change by describing it in a Slack message.
What "65% AI-Written Code" Actually Means
The 65% number is not about AI capability — it is about trust calibration. When Boris says the product team's code is 65% AI-written, he means: 65% of the code that ships has been generated by Claude, reviewed by humans, and judged production-quality. The 35% that remains human-written is likely the architecturally significant code — the core abstractions, the security-critical paths, the novel algorithms. The pattern that emerges: AI handles the well-defined, testable, repetitive code. Humans handle the ambiguous, high-stakes, creative code. That division is not temporary — it is the new equilibrium. And as AI capability rises, the 65% grows not because AI gets smarter but because humans get better at specifying what they want and calibrating their trust.
What To Do
If you are a senior engineer: your value is shifting from "writing code" to "defining what code should be written and judging whether what arrived is correct." Invest in specification skills, review skills, and architectural thinking — not in competing with AI on typing speed. The engineers who thrive in this era are not the fastest coders but the best spec-writers: people who can describe a feature so precisely that AI generates the right thing on the first attempt, and who can spot the subtle bug in AI-generated code that looks clean but hides a logical flaw. If you are a team lead: the 65% number is a target, not a threat. Set up Claude Tag in your team's Slack, define clear approval workflows, and measure how much of your team's routine code can safely move to AI generation. The teams that calibrate this first will ship faster than those still debating whether AI should write code at all. If you are an executive: the three-leap framework gives you a maturity model. Where is your organization? Still at Copilot (autocompletion)? At agent coding (Claude Code)? Or at delegation (Claude Tag in Slack)? The gap between stages is organizational, not technological — it requires workflow redesign, not just tool procurement.