marimo-pair vs SkillClaw

marimo-pair is an open-source extension for marimo reactive notebooks that lets you drop AI agents directly into live, running notebook sessions. Unlike traditional AI coding assistants that only see static code, these agents can execute cells, inspect in-memory variables, read dataframes, manipulate UI components, and iterate on your actual live state — not a static snapshot. The tool plugs into Claude Code via a marketplace plugin and supports any agent implementing the Agent Skills standard. An agent that can see and run your notebook opens up genuinely new workflows: "explore this dataframe and tell me what's anomalous," "run this hypothesis test on the data already in memory," or "generate a chart for each of these 12 conditions." It's the difference between an assistant that reads your code and one that works alongside you in your actual environment. Marimo itself is already a compelling React-based replacement for Jupyter — every cell tracks its dependencies so the notebook is always consistent. marimo-pair makes that reactive model collaborative with AI, enabling a new style of human-AI pair programming where the agent shares your full computational context.

SkillClaw is a research framework from Alibaba's AMAP-ML team that enables collective skill evolution for LLM agent systems deployed at scale. The core idea: instead of each user's agent interactions existing in isolation, SkillClaw aggregates anonymized skill-improvement signals across all users to continuously refine a shared library of reusable agent skills — without requiring centralized fine-tuning. The framework introduces a three-component architecture: a Skill Extractor that identifies and catalogs atomic capabilities from interactions, a Skill Evolver that proposes improvements based on aggregate feedback, and a Skill Selector that routes tasks to the best-available skill version per user context. Published on April 9 and hitting #1 on Hugging Face trending papers this week with 277 upvotes, the paper reports significant improvements over per-user baselines on complex multi-step agentic tasks. This matters especially for production agent deployments where cold-start problems are severe — a new user's agent immediately benefits from millions of prior interactions. It's a fundamentally different model of agent improvement than either fine-tuning (expensive, periodic) or RAG (retrieval-only, no learning).