The theme on GitHub today isn't the AI coding agents themselves — it's everything built around them. Developers are shipping frameworks that make agents more capable, more secure, and easier to supervise when you're running several of them in parallel. The question has shifted from "can this agent write code?" to "how do I manage a fleet of agents that can?"
ECC: A Full Harness Layer for AI Coding Agents
affaan-m/ECC picked up over 2,000 stars today and describes itself as a "harness-native operator system for agentic work." The project bundles 246 skill modules, 61 specialized subagents, a memory persistence layer that converts completed sessions into reusable behavioral rules, and a security component called AgentShield — 102 static analysis rules scanning for secrets, hook injection vulnerabilities, and misconfigured MCP servers. What separates ECC from a typical CLAUDE.md configuration is cross-harness compatibility: the same skill library and memory layer adapts to Claude Code, Cursor, Codex, Zed, and GitHub Copilot through platform-specific config adapters, so your investment in tuning one environment carries over to others.
Anthropic Cybersecurity Skills: 754 Practitioner Workflows for AI Agents
mukul975/Anthropic-Cybersecurity-Skills added over 1,000 stars today with a focused premise: 754 structured cybersecurity skills designed for AI agents to actually execute, not just reference. Each skill encodes a real practitioner workflow — step-by-step, not a generic checklist — and is mapped to five industry frameworks simultaneously: MITRE ATT&CK v18, NIST CSF 2.0, MITRE ATLAS for AI/ML adversarial threats, MITRE D3FEND for defensive countermeasures, and NIST AI RMF for AI risk management. The architecture is designed for token efficiency: agents scan YAML frontmatters at around 30 tokens each to find relevant skills, then load full workflows only when needed. Coverage spans 26 domains including cloud security, threat hunting, threat intelligence, and web application security.
github.com/mukul975/Anthropic-Cybersecurity-Skills
cmux: Terminal UX Built for Running Multiple Agents at Once
manaflow-ai/cmux earned 603 stars today for solving a problem that anyone running concurrent Claude Code sessions will recognize: standard macOS notifications have no context. They all say "Claude is waiting for your input" with no indication of which of your eight panes needs attention, and split-pane tab titles shrink until they're unreadable. cmux is a native macOS terminal built on Ghostty that adds vertical workspace tabs showing git branch, PR status, working directory, and listening ports alongside each session. When an agent needs you, the pane gets a blue notification ring and its tab lights up; a single keystroke jumps to the most recently unread agent. The author frames it deliberately as "a primitive, not a solution" — composable building blocks you wire into your own workflow rather than a rigid opinionated setup.