Claude Code Next-Gen Upgrade and Team Role Restructuring: Background Sub-Agents Become Default, Five New Roles Replace Traditional Positions

Anthropic's AI coding tool Claude Code is getting a major upgrade: sub-agents will run in the background by default, allowing users to chat with Claude while multiple agents execute code refactoring, testing, and PR creation in parallel. Boris Cherny, the creator of Claude Code, also proposed a new team division framework for the AI era, replacing traditional roles (engineer, PM, designer, etc.) with five behavior-based roles: Prototyper, Builder, Sweeper, Grower, and Maintainer.

Next-Gen Upgrade: Background Sub-Agents Run by Default

Boris Cherny announced on X that the next version of Claude Code will enable background sub-agents by default. Users no longer need to manually specify; agents will work in parallel in the background while users continue chatting with Claude. This design upgrades Claude Code from a "question-and-answer dialog box" to a "workflow engine capable of orchestrating multiple task lines."

Previously, Claude Code already had several capabilities:

• Routines: Package prompts, code repositories, and connectors into fixed workflows triggered hourly, nightly, or weekly, or via API, GitHub events, or Webhooks, running on Anthropic's managed cloud.
• Dynamic workflows: Users just mention "use a workflow" or enable ultracode, and Claude generates orchestration scripts to schedule dozens to hundreds of sub-agents in phases, with parallel cross-validation.

This upgrade bundles these capabilities as default behavior, requiring no manual toggling.

Five New Roles: Traditional Job Labels Removed

Boris Cherny proposed on X that as engineering, product, design, and data science functions merge, the Claude Code team no longer uses traditional job titles. Instead, they divide roles based on behavior patterns into five categories:

• Prototyper: Generates many ideas, most of which won't ship; focuses on quantity and disruption.
• Builder: Quickly turns prototypes into production-grade products or infrastructure; core is execution speed and engineering judgment.
• Sweeper: Simplifies UI, streamlines code and system architecture, cuts redundant features, optimizes performance.
• Grower: Takes over mature products and iterates to improve product-market fit (PMF).
• Maintainer: Guards mature systems, ensuring security, reliability, and efficiency.

Cherny emphasized that these roles are not tied to specific positions. Inside Anthropic, designers, engineers, PMs, and data scientists may fall into different roles, and many people span 2-3 roles simultaneously. Roles change dynamically with project phases:

• New products (no PMF): Need many Prototypers, Builders, Sweepers.
• Growth products (initial PMF): Shift focus to Builders, Sweepers, Growers, with a few Maintainers.
• Mature products (strong PMF): Primarily Sweepers, Growers, Maintainers, with some Builders.

Real-World Data: Engineer Output Triples, Bottleneck Shifts to Decision-Makers

Claude Code's effectiveness has been validated both inside and outside Anthropic:

• Inside Anthropic: Claude Code contributed 4% of public commits on GitHub, with annualized revenue exceeding $2.5 billion. Each engineer's output is triple the headcount; a five-person team does the work of 15-20 people. Anthropic's growth team decided to hire more PMs rather than engineers because the bottleneck has shifted from "writing code" to "deciding what code to write."
• Spotify: Engineering VP Niklas Gustavsson revealed that in their monorepo of over 20 million lines of code, there are about 4,500 production deployments daily, 73% of pull requests are AI-assisted, and PR frequency increased by over 75%. He personally runs 5-10 Claude sessions simultaneously, each with an independent git worktree, letting multiple agents work in parallel while he only reviews diffs and makes decisions. Spotify also opens this capability to non-engineers: PMs, designers, and even co-CEOs can describe ideas in natural language, and Claude implements end-to-end prototypes in real code.

Impact and Discussion: Job Boundaries Melt, Golden Age of Generalists

Boris Cherny believes AI is compressing the gap between skills, ushering in a golden age for generalists. He hasn't handwritten code in eight months; all code under his name is 100% generated by Claude. He predicts that in three years, the number of people writing code and using agents will be 100 times today's, but the title "software engineer" may disappear, replaced by more flexible roles.

The community response has been positive. One data scientist said they often act as a Sweeper while also bringing data science taste to product building, a "2+3 type." Senior engineer Kun Chen pointed out that roles should change with projects; locking oneself into one role limits growth. Some questioned: if AI can write code, why do we need Builders and Sweepers? Cherny responded that Claude already handles these tasks well and will improve, but humans still need to take responsibility for AI outputs.

The Sweeper role is particularly underappreciated. As AI-generated code surges, the proportion of AI code going directly into production without human review has risen from about 7% in early 2026 to nearly 38.5% (per Cursor's Developer Habits Report), increasing hidden system issues and highlighting the Sweeper's value in safeguarding quality.