AMUX vs SkillClaw

AMUX is an open-source agent multiplexer that lets you run dozens of Claude Code (or other terminal AI coding agents) simultaneously, all managed from a single web dashboard — no complicated setup required. Built by the team at Mixpeek, it requires only Python 3 and tmux, with the entire server delivered as a single ~23,000-line Python file with embedded HTML/CSS/JS. The standout features are a self-healing watchdog that auto-compacts context when it drops below 20% and restarts stuck sessions, a SQLite-backed kanban board where agents atomically claim tasks to prevent duplicate work, and a REST API injected at startup that allows agents to coordinate with each other via simple curl calls. There's even a mobile PWA with offline support via Background Sync so you can monitor your agent army from your phone. In the "agentmaxxing" era, AMUX is the most complete open-source solution for running parallel AI coding agents unattended. Rather than babysitting one agent, you dispatch 5–20 agents to isolated worktrees and check back in as a reviewer. The MIT + Commons Clause license means it's free to self-host.

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).