Boris Cherny on coding from his phone and the post-coding bottleneck
Key takeaways from the Scale fireside chat with Anthropic Head of Claude Code Boris ChernyDealroom has a profile for this one. Try Dealroom → (June 2026):
100% AI-written code. Cherny says every line of his code has been written by Claude Code since Opus 4.5 launched in November — he uninstalled his IDE entirely. This year he shipped roughly 1,700 pull requests, adding 400,000 lines and deleting 250,000, and has used over 8 billion tokens since March.
Coding from a phone. Most of his coding is now done on his phone rather than a laptop — a shift he says he would have called crazy six months ago, driven by prompting (often by voice) replacing hand-written code.
Optimise for ROI, not cost. His advice to companies: give everyone tokens — not just engineers, but PMs, designers, data scientists — let teams experiment, then control costs on the back end once use cases take off. He argues there is maybe a 50% opportunity to cut investment but a far larger opportunity to grow returns, so focus on upside.
An 8x jump in code per engineer. Anthropic has seen a roughly 8x increase in code per engineer since the start of the year. Once a team hits 100% AI-written code, the bottleneck shifts upstream to good ideas (more PMs, user researchers) and downstream to go-to-market and marketing.
Loops as the next abstraction. Cherny frames "loops" — agents prompting agents that write code — as the next step up the abstraction ladder, as big a leap as source code to agents. He estimates roughly 30% of his code is now written by loops on an average day.
Co-work for non-engineers. Anthropic's Co-work (in the Claude desktop app) is "Claude Code for non-engineers" built on the same agent SDK, with extra guardrails. He uses it for project management, status round-ups in Slack, and fully automated travel booking that scans his calendar and books flights and hotels.
Solving the next bottleneck. With coding largely solved, Anthropic built Claude Code Review (catching ~98-99% of bugs before a human sees the PR) and Claude security review (now catching issues pen-testers miss). Across Anthropic, 80-90% of code is written by Claude, with many teams at 100%.
Auto mode beats prompt fatigue. Constant yes/no permission prompts caused engineers to rubber-stamp approvals, hurting security. Anthropic's new auto mode routes each permission decision to a model, which it says is more secure than both default and dangerous modes — and unlocks agents running for hours or days.
Source: Scale fireside chat hosted by a Meta director of product — June 2026.