Agentic Research

May 2026 W2 AI Roundup: Skills Ecosystem Explodes, Terminal Open-Source Wave, Enterprise AI Arms Race

2026/05/1125 min readBryan Chan閱讀中文原文
TopicsWeeklyGitHubAgentDeepSeekOpen Source

This week's brief is based on a DuckDuckGo editorial summary, supplemented by cross-verification with Tavily + Brave. For the search strategy, see: Three-Engine Search Strategy


1. GitHub Trending: The Skills Ecosystem Tops the Chart

This week (4/28 - 5/6), GitHub's trending list was dominated by a single theme: Agent Skills.

🥇 mattpocock/skills: A TypeScript Legend Enters (+31K Stars)

Matt Pocock, the most authoritative technical educator in the TypeScript community, open-sourced a set of Agent skills: 17 structured SKILL.md files covering:

  • TDD (test-driven development)
  • Triage (issue classification and prioritization)
  • ADR (architecture decision records)
  • Structured code review

Unlike Addy Osmani's agent-skills, Matt Pocock's skills lean more toward day-to-day development workflows (issue management, documentation generation), while Addy's lean toward software engineering best practices (Spec-Driven, five-axis Review). The two are complementary and together form a complete developer skill ecosystem.

🥈 awesome-codex-skills: The Ecosystem Reaches 1000+ (+4K Stars)

The Codex skill directory curated by ComposioHQ passed 1000+ SKILL.md packages this week. This means Agent Skills has moved from an experiment by a few early movers into a mature field with a directory, standards, and an ecosystem.

GitHub Itself Enters: Spec-Kit

On May 8, GitHub released Spec-Kit, an open-source Spec-Driven Development toolkit. This is the first time GitHub has officially recognized that "write the spec before the code" is the standard workflow of the AI era.


2. The Terminal Open-Source Wave: Warp & DeepSeek-TUI

Warp Goes Fully Open Source (AGPL-3.0, +37K Stars)

The AI terminal Warp, written in Rust, announced it is fully open source under the AGPL-3.0 license. Warp is no longer just "a better terminal"; it has repositioned itself as an Agentic Development Environment, a development environment with a built-in AI Agent. It is used by 700,000+ developers and is backed by OpenAI investment.

DeepSeek-TUI: A Terminal Built for DeepSeek V4

A terminal coding Agent designed specifically for DeepSeek V4, integrating the DeepSeek API with a terminal UI so developers can interact with DeepSeek V4 directly from the command line.


3. DeepClaude: Claude Code × DeepSeek V4 Pro, Costs Cut 17x

The project that drew the most attention on Hacker News this week (669 points). DeepClaude is an open-source bridge layer that lets Claude Code's Agent Loop (task decomposition, autonomous execution, self-correction) run on a DeepSeek V4 Pro backend.

Why does this matter?

Claude APIDeepSeek V4 Pro
Input / 1M tokens$15$0.55
Output / 1M tokens$75$2.19
Cost multipleBaseline17-34x cheaper

Real-world testing shows that on code generation and architecture analysis tasks, DeepSeek V4 Pro performs on par with Claude or even better, at only 1/17 the cost.


4. Enterprise-Grade Agent Skills Arrive: Kaltura Goes Open Source

On May 4, NASDAQ-listed company Kaltura (KLTR) announced it was open-sourcing a set of AI Agent Skills. This is the first publicly listed company to enter the Agent Skills race.

Kaltura's skills are designed for Claude Code, OpenAI Codex, GitHub Copilot, Cursor, and others, with these characteristics:

  • Machine-readable: not traditional documentation, but structured knowledge modules
  • Production-validated: already battle-tested on the Kaltura platform
  • Rich-media focused: covering video portals, interactive webinars, virtual avatars, and other scenarios

This marks the move of Agent Skills from individual developers (Addy Osmani, Matt Pocock) into an enterprise-grade phase.


5. The Model Arms Race

ModelRelease TimeKey Highlights
GPT-5.54/23Brand-new foundational architecture, top of Terminal-Bench 2.0, multimodal capabilities
Claude Opus 4.7AprilReclaims the SWE-Bench coding crown
Gemini 3.1 UltraApril2M-token context window, native Mac app
GLM-5.1March-AprilMajor update to the Tsinghua-affiliated open-source model
Gemma 4March-AprilGoogle's open-source lightweight model
DeepSeek V4RecentlyFlash $0.14/1M + Pro 1M context

6. The Enterprise AI Arms Race

OpenAI $10B vs Anthropic $1.5B: Announced the Same Day

On May 4, the two AI giants announced joint ventures with private equity on the same day:

OpenAIAnthropic
Scale$4B raise, $10B valuation$1.5B
PartnersTPG, Brookfield, Advent, BainBlackstone, Goldman Sachs, H&F
PositioningLarge-enterprise AI deploymentClaude rollout for mid-sized enterprises
Core bottleneckBoth pointed to the same problem: a shortage of engineers who can implement AINot applicable

This shows that competition in the AI industry has shifted from "whose model is better" to "who can actually deploy AI inside enterprises".

Microsoft Agent 365 Reaches Full GA

Microsoft announced the full availability of Agent 365, with agentic workflows integrated into Word, Excel, PPT, and Outlook. It also released the 2026 Work Trend Index report: organizational culture has more than twice the influence on AI adoption than individual factors.

Cloudflare Lays Off 1,100, Pivots Fully to Agentic AI

Cloudflare announced layoffs of 20% (1,100+ people). CEO Matthew Prince said: "Internal AI usage grew 600% in three months, and many human workflows have become redundant." Following Oracle, Microsoft, and GoPro, this is another tech company carrying out large-scale layoffs citing AI.


7. OpenClaw: 280K Stars, Surpassing React

OpenClaw passed 280,000 Stars on GitHub, officially surpassing React to become the world's most popular AI Agent framework. In Reddit community discussions, OpenClaw is repeatedly recommended as the leading open-source option for an "always-on persistent agent".

  • Supports 20+ messaging platforms (Telegram, WhatsApp, Discord, Slack, and more)
  • Built-in cron scheduled tasks, heartbeat monitoring, and multi-Agent routing
  • The Skill system lets users extend functionality without writing code

8. Panorama of the Skills Ecosystem

By May 2026, Agent Skills has formed a complete ecosystem:

Individual developer tier      Addy Osmani (21 skills)  ← Software engineering best practices
Matt Pocock (17 skills)   ← Daily development workflow

Enterprise tier          Kaltura (NASDAQ-listed)     ← Production-proven

Catalog and standards layer      awesome-codex-skills      ← 1000+ package curation
GitHub Spec-Kit           ← Official Spec-Driven Standard

Framework Layer   OpenClaw Skills           ← Top 100 community contributions
                  Claude Code Skills        ← Native support

Recommended Reading