Why “Claw” Is the New Layer of AI Agents and What It Means for Software Development

1. Karpathy Introduces a New Term

Andrej Karpathy, a leading AI figure, coined the term "Claw" to describe a new layer on top of large language model (LLM) agents. He bought a Mac mini to experiment with OpenClaw, an open‑source AI‑agent project that has amassed 200 k GitHub stars.

Karpathy warns that OpenClaw’s 400 k lines of code are largely generated by AI "vibe coding" and have been subject to large‑scale attacks, with malicious code infiltrating plugin markets.

He concludes that Claw is not a replacement for agents but a foundational infrastructure that lets agents "stay alive".

2. What Exactly Is Claw?

Karpathy uses a concise table to compare three layers:

Large Model (e.g., ChatGPT) – acts as a consultant, responding only when asked.

AI Agent – serves as an executor that can decompose tasks, call tools, and go offline after completion.

Claw – a resident employee with persistent memory, scheduling, message routing, and a local runtime that can access files and LAN devices.

Claw adds the following capabilities on top of agents:

A continuously running process.

A scheduling mechanism for timed tasks.

Unified message routing for various chat platforms.

Persistent cross‑session memory.

A local runtime bound to specific hardware.

3. Why Is This Emerging Now?

In two weeks, OpenClaw’s token usage surged to about 13 % of all tokens on OpenRouter, and its GitHub stars jumped from zero to 200 k within months.

Competing lightweight forks such as NanoClaw (1 % of OpenClaw’s code, viewable in eight minutes), ZeroClaw, PicoClaw, and several Chinese variants are appearing rapidly.

Simon Willison notes that "Claw" is becoming an industry‑wide term for architectures like OpenClaw.

4. What Does Karpathy Value?

He highlights two strengths of NanoClaw:

Only ~35 k tokens (≈17 % of Claude Code’s context window), allowing the entire codebase to be read and understood in one pass.

Use of /add-telegram “skills” instead of static configuration files, letting the AI modify code automatically based on natural‑language instructions.

This approach replaces manual config edits with AI‑driven code changes, embodying the insight: "Write a maximally adaptable code base, then use skills to shape it into any desired form."

5. Deeper Implications

Karpathy likens this to the 2017 MAML paper, which trains a model to adapt quickly to new tasks with few examples. Similarly, NanoClaw serves as a software‑level MAML: a fork‑friendly repository that can be rapidly repurposed.

The key takeaway is that the most valuable software is not the most feature‑rich, but the easiest to transform.

6. Rethinking Classic Software Principles

Gavriel Cohen, NanoClaw’s author, argues that traditional DRY, strict file‑size limits, and code‑beauty standards are outdated when AI writes and maintains code. Copying code for safety, allowing larger files, and focusing on functional adequacy become preferable.

"We don’t need today’s code to survive for years; it only needs to work now, because stronger AI will rewrite it later."

Thus, when AI dominates code creation and maintenance, software engineering principles must be recalibrated.

7. Why Run Claw Locally on a Mac Mini?

Karpathy emphasizes three reasons for local deployment:

Privacy – emails, calendars, and files stay on the device.

Latency – zero‑delay local calls enable sub‑second AI responses.

Control – owning the hardware means owning the AI.

The Mac mini has become the ideal "physical embodiment" for Claw, explaining its recent sales surge.

8. Final Thoughts

Karpathy’s tweet packs dense insights: a new term, a new layer, and a new direction for AI‑augmented software development. He demonstrates that top‑tier technologists learn by hands‑on experimentation, embracing imperfection and prioritizing adaptability over exhaustive feature sets.