The midfield battle of coding: agents cross the Rubicon

Crossing the Rubicon

It has been reported that a decisive shift occurred this May: large-language-model makers collectively crossed the boundary between “assistant tools” and “primary productivity agents” in software engineering. Anthropic’s Claude Code — an autonomous coding agent — is central to that leap, and the commercial numbers are jaw‑dropping. It has been reported that Anthropic’s annualized revenue jumped from roughly $10 billion to $44 billion in three months, with daily gains near $96 million. Who still thinks coding is just autocomplete?

How Anthropic won the lane

The strategy that reportedly made the difference was simple and contrarian: train on messy, real‑world codebases rather than sanitized contest problems. It has been reported that Anthropic co‑founder Jared Kaplan argued to feed models “dirty” repositories so agents learn to handle legacy systems, brittle dependencies and real engineering debt. The payoff arrived in stages — Sonnet 3.5 built project‑level context, Sonnet 3.7 added autonomous terminal actions, and Claude Code turned into a production agent that some customers say performs the bulk of day‑to‑day coding. Reportedly Claude Code already accounts for a dominant share of AI programming traffic and has driven rapid enterprise adoption from fintechs to e‑commerce platforms.

The global scramble

The result: every major AI cloud and model maker is rushing across the same river. OpenAI has tried to recover from earlier strategic missteps around Codex; GPT‑5.5 plus a /goal mode produced striking lab results and OpenAI briefly offered Codex free to jump‑start adoption. Google’s play is price and openness — Gemini CLI and new enterprise tooling aim to make agents cheap and embedded in Google’s cloud stack. Meta, late to the fight, has also been aggressive; it has been reported that Mark Zuckerberg personally contributed code using a competitor’s agent. This is no longer a feature war. It is a contest for developer mindshare, cloud lock‑in and the pathway to broader AGI capability.

Stakes beyond engineering

This race unfolds amid broader geopolitics — US‑led export controls on advanced chips, national AI strategies and cloud competition all shape who can deploy these agents at scale and where. Chinese technology groups such as Baidu (百度), Alibaba (阿里巴巴) and Tencent (腾讯) are watching closely: control of coding agents determines who owns developer platforms, cloud revenue and, ultimately, critical soft‑infrastructure. It has been reported that firms now see the move from “assistant” to “productivity主体” as irreversible. The Rubicon has been crossed — coding is the frontline, and everyone has taken sides.