Andrej Karpathy talks about "Claws"
AI pioneer identifies 'Claws' as a new orchestration layer above LLM agents, sparking a wave of lightweight implementations.
In a blog post by Simon Willison, AI pioneer Andrej Karpathy discussed the emerging concept of 'Claws'—a new layer in the AI stack built on top of LLM agents. Karpathy, known for coining terms like 'vibe coding,' argues that just as agents were a layer on LLMs, Claws are now a layer on agents, handling advanced orchestration, scheduling, context, and persistent tool calls. He specifically mentioned experimenting with OpenClaw and highlighted lightweight implementations like NanoClaw, whose core engine is roughly 4000 lines of code, making it manageable and auditable. The trend includes systems like nanobot and zeroclaw, designed to run on personal hardware and communicate via messaging protocols. Karpathy's endorsement signals that 'Claw' is becoming a recognized term for this category of personal, orchestrated agent systems.
- Andrej Karpathy identifies 'Claws' as a new orchestration layer above existing LLM agents, handling scheduling and persistent tool calls.
- Lightweight implementations are emerging, like NanoClaw with a ~4000-line core engine designed to be auditable and run in containers.
- The concept signifies a shift toward personal, hardware-based agent systems that can act on instructions and schedule tasks autonomously.
Why It Matters
Defines the next evolution of AI agents, moving orchestration and persistence to personal hardware for more autonomous, scalable workflows.