From 'Vibe Coding' to 'Claws': Two AI Masters Map a Fast Shift from Feel to Fully Agentic Work

Karpathy names the feeling, then accelerates the story

Andrej Karpathy has done it again. The AI engineer and commentator coined "Vibe Coding" to describe code written in a state of flow — and it has been reported that Collins picked the term as word of the year within nine months, an unusually quick ascent that signalled something deeper than a catchy tweet. Karpathy’s new coinage, "Claws", extends that pattern: not just a label but a proposition. He describes Claws as a new layer above LLM-based agents that handles orchestration, context, tool-calling and persistence — and crucially, that this layer can run on personal hardware such as the Mac Mini. Reportedly, demand for small local machines to run agents has surged, underscoring the move from editing code with AI help to delegating entire workflows to AI.

Willison builds the how

If Karpathy sketches the horizon, Simon Willison builds the ladder. The Django co‑creator has shifted from naming vulnerabilities like "prompt injection" to publishing practicable work: his Agentic Engineering Patterns project opens with the blunt claim "Writing code is cheap now." That framing reframes the problem. If code can be generated and composed cheaply, the harder work becomes system design, safe orchestration and durable state — the exact spaces Karpathy’s Claws targets. Willison’s playbook is aimed at practitioners: patterns, testing approaches and constraints that turn speculative agent hype into repeatable engineering.

A practical inflection with broad geopolitical reverberations

Why does this matter beyond Silicon Valley? For developers and product teams in China — and for policy watchers in Washington and Beijing — the shift from human-crafted apps to agent-managed services changes technology stacks and supply chains. It has been reported that Chinese cloud and AI vendors, including Baidu (百度), are accelerating support for agent frameworks and lightweight local inference precisely because export controls on high-end chips make onshore, efficient alternatives more attractive. The result: more logic moves into orchestration layers and smaller devices, not just bigger models in distant clouds.

What ordinary users and governments should watch

So what happens next? Ordinary users may never learn to code — and they won’t need to. They will interact with assistants that stitch together tools and ephemeral apps to meet requests. Regulators, meanwhile, must catch up: who audits an agent that composes, tests and deploys ephemeral code? Karpathy and Willison are offering complementary visions — one naming where we’re headed, the other showing how to get there — and their combined influence is already reshaping how engineers, companies and states think about software, control and trust.