article

AI Dev Essentials #30: Cursor 2.0, GitHub Agent HQ, and Cognition SWE-1.5

AI dev news: GitHub Agent HQ, Cursor 2.0, Cognition SWE-1.5, and OpenAI's $1.4T plan. Plus, get a new 6-lesson course on building Claude Code skills.

Hey Everyone 👋,

John Lindquist here with the 30th issue of AI Dev Essentials!

I was invited to attend GitHub Universe this week. They covered my conference and travel costs (thanks Ashley and Christina!). It was packed with AI/Agent announcements. The Agent HQ announcement positions GitHub as the central hub for any AI agent (Claude, Codex, Jules, Grok, etc) and they're all available through single subscription with Copilot Pro+. Cursor launched version 2.0 with their first proprietary Composer model and parallel execution. Cognition shipped SWE-1.5 at 950 tokens per second thanks to their partnership with Cerebras.

I've been focusing on building out Claude Code skills and plugins. In fact, I published a plugin for the programmable badge they handed out at GitHub Universe (https://github.com/johnlindquist/badger-2350-plugin). It was a good exercise in preparing a plugin w/ skills for working in a completely new environment to enable anyone to build apps without prior experience.

I'll be covering more about skills and plugins in my upcoming workshop. Sign up here: https://egghead.io/workshop/claude-code

🚨 Early Bird Pricing for the Workshop Ends Tomorrow! 🚨

🎓 New egghead.io Lessons This Week

I've been releasing a steady stream of new lessons covering all the aspects of Claude Code skills:

The Essential Guide to Claude Code Skills - 6 Lessons Published

I've been focused on launching a comprehensive new course teaching developers how to build custom skills for Claude Code. All six lessons are now live and available free as a community resource. Here's what we cover:

Lesson 1: Create Your First Claude Code Skill Learn the core concepts of skill development by building a timestamp skill from scratch. Understand the required file structure, YAML frontmatter, and how to safely grant Claude access to specific shell commands.

Lesson 2: Control Claude Skills Output with References and Examples Master precision in AI agent behavior by using detailed example files. Learn how to enforce specific file naming conventions and folder structures while keeping context usage efficient through lazy-loaded examples.

Lesson 3: Stacking Claude Skills to Create Complex Workflows Discover how to build modular, composable skills that work together. Learn to define skill dependencies so Claude can autonomously orchestrate multi-step workflows by chaining multiple skills.

Lesson 4: Build Better Tools in Claude Skills with Scripts Escape skill limitations by abstracting complex logic into TypeScript/Bun scripts. Learn how to bypass parser issues with shell redirects and achieve cross-platform compatibility through programmatic control.

Lesson 5: Secure Your Claude Skills with Custom PreToolUse Hooks Implement fine-grained security controls using Claude Code hooks. Learn to programmatically validate commands and enforce strict security policies before tool execution.

Lesson 6: Claude Skills Compared to Slash Commands Understand the key differences between Slash Commands (user-invoked) and Skills (agent-driven). Learn when to use each pattern to create more efficient and powerful AI-assisted workflows.

Start the course →

🚀 Major Announcements

GitHub Universe 2025: Agent HQ Unifies AI Coding Agents

GitHub announced Agent HQ on October 28, 2025, at GitHub Universe 2025, creating a unified platform to orchestrate any AI agent from any provider through GitHub Copilot subscriptions.

Platform capabilities:

This represents GitHub's strategic move to become the central hub for AI coding, regardless of which underlying agent provider developers prefer.

(GitHub Blog: Agent HQ, GitHub Changelog: Custom Agents, VentureBeat, CNBC)

I've had a lot of conversations around background agents vs. focused work recently. Everyone is trying to find that line of "when can trust an Agent to complete a task without help?". The answer is different for every task, every model, and every user. So it's extremely difficult to generalize the marketing the messaging for tools like Agent HQ. Often you spend so much time configuring the Agents that you could have just written the code yourself. Other times you're configuring an Agent that you'll never use again. It's becoming a delicate balancing act. There's also a strong argument that you need to set up your Agents now so that they're ready once the models improve enough to trust them. Regardless, these are important skills to have in your toolbelt and I strongly recommend putting in the work to get comfortable with them.

Cursor 2.0 Launches with Composer Model and Multi-Agent Orchestration

Cursor released version 2.0 on October 29, 2025, introducing Composer, their first proprietary coding model, along with infrastructure for running up to 8 agents in parallel on a single task.

Technical breakthroughs:

(Cursor Blog: 2.0 Release, InfoWorld, VentureBeat, Bloomberg)

If you'd been using the "Cheetah" model the past month, that was an early preview of Cursor's latest Composer model. It's wonderfully fast and smart enough for most grunt work. I don't quite trust it nearly as much as I trust Sonnet 4.5 to fully work through and complete a task, but the speed offers up an tight feedback loop that helps you stay focused on a single task. I'll need to spend way more time with it to get a better feel for when I want to use it. I definitely trust it for simple chores and it's amazing watching it knock out a task in seconds.

Cognition Releases SWE-1.5: Near-SOTA Performance at 950 Tokens/Second

Cognition announced SWE-1.5 on October 29, 2025, a frontier-size agent model achieving near state-of-the-art coding performance while delivering unprecedented speed through partnership with Cerebras.

Performance highlights:

The release emphasizes Cognition's thesis that the future requires vertical integration across the entire stack, not just model improvements.

(Cognition Blog: SWE-1.5, Testing Catalog)

Not to be outdone, Windsurf released their own model later in the same day. The speed comes from a partnership with Cerebras and their custom chips. I'd be curious how fast the model would be on the same hardware as Cursor's, but at the end of the day it doesn't really matter to us end-users. As long as they can keep the prices down, I'm all for competition. Honestly, I need to re-install Windsurf and try it out again (even though I'm extremely satisfied with Cursor)

Sam Altman Outlines OpenAI's Trillion-Dollar Vision for Automated AI Researchers

OpenAI CEO Sam Altman announced on October 28, 2025, ambitious timelines for automated AI research capabilities, alongside massive infrastructure commitments totaling $1.4 trillion in financial obligations.

Strategic roadmap:

Altman emphasized this represents a calculated bet on future model capabilities and revenue growth justifying the unprecedented infrastructure investment.

(TechCrunch, TechRadar, Technology.org, Business Standard)

$1.4 trillion is an absurd amount of money to bet on a specific vision of how AI scales. If I was to spend that kind of money on something, I'd certainly hope it was curing cancer and solving climate change. The Automated AI Research Interns sounds like an awesome project and I'm really cheering for them to succeed. Not because I particularly care about OpenAI, I just want the world to be a better place.

OpenAI Releases gpt-oss-safeguard for Customizable AI Safety

OpenAI launched gpt-oss-safeguard on October 29, 2025, an open-weight safety toolkit enabling developers to build customizable content moderation systems using reasoning-based policy interpretation.

Safety innovation:

The release represents OpenAI's strategy to open-source safety infrastructure while keeping frontier models proprietary.

(OpenAI Blog: gpt-oss-safeguard, OpenAI Technical Report, WinBuzzer, CNBC)

🛠️ Developer Tooling Updates

VS Code Integrates OpenAI Codex Through Agent Sessions

Visual Studio Code announced OpenAI Codex integration on October 28, 2025, bringing cloud-based AI coding agents directly into the editor through the new Agent Sessions view.

Integration features:

The integration represents Microsoft's strategy to bring multi-agent orchestration directly into developers' primary workflow tools.

(GitHub Changelog: VS Code Upgrade, Visual Studio Code Release Notes, Visual Studio Magazine)

I honestly think this is Microsoft's game to lose. They have GitHub and VS Code. They have all the users and all the infrastructure. So it comes down to the UX and their partnerships with the model providers. I'm honestly a little surprised they're not announcing models of their own yet. From all of the various Microsoft + OpenAI partnerships, I suspect we'll see tighter integrations between their products moving forward, especially because OpenAI doesn't seem interested in forking VS Code like everyone else.

Stitch and Jules Integration Streamlines Design-to-Code Workflow

Google's Stitch team announced Jules integration in late October 2025, enabling seamless transitions from design mockups to working code in multiple frameworks.

Workflow capabilities:

The integration represents Google's strategy to connect design tools with agentic code generation workflows.

(Google Developers Blog: Stitch Launch, Medium: Stitch and Gemini Integration)

Gemini is such an awesome model for design. And while I'd still consider these project "Labs", they certainly have my attention and I can't wait to see what I can build with it.

Vercel Partners with Z.ai for Lowest-Cost GLM 4.6 Access

Vercel announced partnership with Z.ai in late October 2025, offering GLM 4.6 through AI Gateway at highly competitive pricing.

Partnership details:

The partnership represents Vercel's strategy to offer diverse model options through their infrastructure layer.

(Vercel AI Gateway: GLM 4.6, Z.ai Blog: GLM-4.6)

Google Launches Pomelli Marketing Agent on Google Labs

Google released Pomelli on October 28, 2025, a specialized AI agent for marketing campaign generation, available through Google Labs in select regions.

Agent capabilities:

The launch represents Google's strategy to create specialized agents for specific business functions beyond general-purpose AI assistants.

(Google Blog: Pomelli, Search Engine Journal)

This is a neat proof of concept. My first impression we're "fine", it's not something I would use today, but it has a ton of potential for the future.

🤖 AI Ecosystem Updates

Gemini CLI Adds Interactive Shell and Tool Calling

Google released Gemini CLI versions 0.9.0 and 0.10.0 in early October 2025, introducing interactive shell capabilities and intelligent tool usage without explicit commands.

CLI enhancements:

The updates represent Google's commitment to making Gemini CLI a full-featured development environment rather than just a chat interface.

(GitHub: Gemini CLI Releases, Gemini API Docs: Release Notes)

Addy Osmani Publishes Comprehensive Gemini CLI Tips Guide

Chrome engineering manager Addy Osmani published an extensive Gemini CLI tips and tricks guide on October 21, 2025, covering 30+ pro-level techniques for maximizing CLI productivity.

Guide coverage:

The guide represents community-driven documentation helping developers level up their CLI usage.

(Addy Osmani Substack: Gemini CLI Tips, GitHub: Gemini CLI Tips Repository)

⚡ Quick Updates

Cursor 2.0 Ships Browser for Agent

(Cursor 2.0 Changelog)

OpenAI Sora Character Cameos Expand to Pets and Objects

(OpenAI: Sora 2, TechCrunch: Sora Update)

GitHub Universe 2025 Featured Developer Community

(GitHub Universe Keynote Recap)

Gemini CLI Hit 1 Million+ Developers

(Google Blog: Gemini CLI Extensions)

✨ Workshop Spotlight (🚨 Early Bird Pricing Ends Tomorrow! 🚨)

Claude Code Power User Workshop - November 7th

Date: November 7, 2025 Time: 9:00 AM - 2:00 PM (PDT) Platform: Zoom

Pricing:

What You'll Learn:

Master the essential skills to ship reliable AI-generated code with confidence. This hands-on workshop covers everything from foundational prompting to advanced automation using the Claude Code SDK and custom integrations.

Core Skills:

Register: https://egghead.io/workshop/claude-code

Read this far? Share "AI Dev Essentials" with a friend! - https://egghead.io/newsletters/ai-dev-essentials

https://egghead.io