Claude Code Founder Commands Thousands of AI Agents from a Phone—No Code Needed

Farewell to Manual Coding

In 2024 the state‑of‑the‑art AI coding assistance was "type‑ahead" autocomplete. Anthropic Labs tried to let agents write all code; early versions were poor, barely usable for limited tasks and received lukewarm market response.

The breakthrough came in May 2025 with the release of Opus 4, followed by rapid iterations to 4.5, 4.6 and now 4.7, which caused an exponential rise in model understanding and autonomy. The team’s original belief—that pre‑building product frameworks would let the next‑generation model deliver a qualitative experience—was validated.

Boris Cherny now says the model writes 100 % of his code, and he submits dozens of pull‑requests daily, once even 150 in a single day. Very large or obscure codebases sometimes still trip the agents, but engineers simply wait for the next model release.

Mobile‑Based Automation Empire

Without a keyboard, Cherny runs his workflow from a phone, launching five to ten concurrent Claude sessions that each host hundreds of parallel agents. These agents execute continuous‑integration fixes, system clean‑ups, and periodic data‑collection tasks via a simple loop command similar to cron.

Loop‑driven tasks run even when the phone is off; they handle error‑repair, test‑fluctuation cleanup, and automated user‑feedback classification sent to office tools. The workflow reshapes team roles: every engineer, product manager, designer, data scientist, finance person and researcher now writes programs in natural language, while their domain expertise remains the differentiator.

Software’s Printing Press

The interview draws a parallel between the 15th‑century printing press, which lowered the cost of books and raised literacy, and today’s AI‑driven code generation that lowers the barrier to software creation. As a result, programming becomes a ubiquitous skill comparable to sending a text message.

Examples include a shop owner building an inventory system and a hobbyist programming a micro‑controller to turn on lights on door opening.

Re‑Engineering the Moat

AI erodes traditional software‑service moats. Conversion‑cost barriers disappear because a model can migrate legacy data and logic with a simple instruction. Process‑level barriers also collapse; a 4.7‑level model can iteratively discover and execute an entire business workflow.

Network effects, economies of scale and resource monopolies remain strong, but for startups the AI era offers unprecedented opportunities to build products that rival large tech firms with small teams.

Future Outlook

All database queries and low‑level architecture are now generated by the model. Agents even negotiate integration issues via chat in office software. Over time, harnesses and safety guardrails become redundant as models evaluate task complexity and decide whether to run locally in a container or offload to cloud compute.

Product design is shifting toward enabling massive parallel model computation, making the debate over cloud versus on‑device models moot. Ultimately, the software world may contain no handwritten code, only precise, continuously running ideas in the cloud.