Claude Code and the Developer's New Job: Less Writing Code, More Running the System

If you've been coding with an agent like Claude Code for a while, you've probably felt it: you spend less time typing code and more time deciding what to build, reviewing what the agent produced, and keeping the whole thing from rotting. Boris Cherny, who leads Claude Code at Anthropic, recently put a useful frame on this shift. As "engineering, product, design, DS, etc. melt into a new kind of role," he mapped five archetypes of work on his team — and for a solo developer or a small team, they're less like job titles and more like modes you switch between during a project.

The developer's job is moving up the stack

The mechanical change is simple: when an agent can write most of the code, the scarce work is no longer producing code but directing and governing it. Deciding what to build, specifying it well, reviewing the diff, catching the edge case the agent missed, keeping the system simple, and maintaining it in production — that's the job now. Cherny's five archetypes are a clean checklist for the modes that work splits into.

Why you should think in modes, not titles

Cherny's key observation is that these archetypes aren't tied to a function — on his team, some designers act as prototypers, some engineers as sweepers, and so on. For a solo builder that's liberating: you stop asking "am I a developer or a product person?" and start asking "which mode does this project need right now?" Early on you're a prototyper; once something clicks you switch to builder; before launch you sweep; after launch you grow; and eventually you maintain. The skill that matters is recognizing the transition and changing gears.

Cursor vs Claude Code vs Codex: Which AI Coding Tool Is Worth It?

Practical: how the modes change your Claude Code workflow

• Prototyper mode — let the agent run loose. Vague prompts, throwaway branches, no tests. The goal is signal, not quality; kill most of what you make.
• Builder mode — tighten the spec. Write a clear brief, define the minimal paid workflow, and review the agent's output like a PR. This is where a CLAUDE.md / agent-rules file earns its keep, because you're now optimizing for production shape, not speed of exploration.
• Sweeper mode — point the agent at cleanup: simplify, delete, add tests, fix edge cases, tune performance. Un-shipping is a feature.
• Grower mode — the work moves outside the codebase: onboarding, activation, SEO, benchmarks, social proof. The agent helps, but the leverage is distribution.
• Maintainer mode — automate the boring parts: monitoring, billing, admin, dependency hygiene. Keep the system understandable and cheap.

The mistake is running one mode when the project needs another — sweeping when you should still be prototyping, or prototyping a product that actually needs a maintainer.

FAQ

Does this mean developers write less code?

Increasingly, yes — the typing is delegated to the agent. The developer's value shifts to specifying, reviewing, simplifying, and maintaining. You're running the system more than authoring every line of it.

Are prototyper/builder/sweeper/grower/maintainer real roles?

They're archetypes of work, per Boris Cherny, not job titles. For a solo dev they work best as modes you deliberately switch between across a project's life.

Which mode should I default to?

None permanently. Match the mode to the stage: prototyper to find signal, builder to ship the minimal version, sweeper before launch, grower after, maintainer once it's mature.

Bottom line

Agentic coding tools don't just change how you write code — they change what your job is. Boris Cherny's prototyper–builder–sweeper–grower–maintainer framing is a practical way to run that new job: stop identifying with a title, and start switching modes as the project moves from idea to production to growth to maintenance. The developers who get the most out of Claude Code are the ones who know which mode they're in — and when to change.

Sources and further reading