AI Dev Essentials #20: Claude on GitHub, AGENTS.md & GPT-5 Updates

Hey Everyone 👋,

John Lindquist here with the 20th issue of AI Dev Essentials!

This past week, I've been digging in to customizing Claude Code GitHub Actions. Claude Code now officially supports using your subscription to generate an API key which allows you to use it on GitHub for cron jobs, pull requests, and anything GitHub Actions supports. This is a huge unlock for both cost savings and building your own code maintenance, review agents, etc. Excited to share more about this in the future.

I've been keeping a close eye on OpenAI's Codex CLI with the recent release of GPT-5. They've been making incredible progress and just last night they'll enabled configuring the reasoning effort. Add model_reasoning_effort = "high" to your ~/.config/codex/config.toml file to maximize every codex session :)

I'm really, really enjoying following the reasoning logic of Codex and watching its thinking progress. It's arguably the best reasoning experience I've seen in a CLI and I'm seriously tempted to use it as the starting point for all of my coding tasks then piping the plans into Claude Code to execute... I'll update you next week on that experience.

Lastly, I also recorded and published five lessons focusing on building custom agents in Claude Code. This is an area I'm extremely passionate about and is a huge unlock for overcoming the "stuck" feeling you sometimes get when starting a new session. I'm excited to cover all of this and more in my upcoming workshop: https://egghead.io/workshop/claude-code

🎓 New egghead.io Lessons This Week

Automating Development Workflows with Claude Code Agents

A collection of practical recipes for automating your development workflow with custom-built Claude Code agents. This course provides a hands-on guide where you'll learn to craft precise system prompts to create reliable, single-purpose agents, package them as simple command-line tools using bun build, and grant them the autonomy to complete multi-step tasks like web research and file creation.

5 lessons covering:

Crafting precise system prompts for single-purpose agents
Packaging agents as command-line tools with bun build
Building autonomous multi-step task workflows
Web research automation strategies
File creation and management patterns

Check out the course here: https://egghead.io/courses/claude-code-automation-cookbook-recipes-for-ai-agents~tsilh

🚀 Major Announcements

AGENTS.md: The New Standard for AI Agent Context

The AI coding ecosystem has coalesced around AGENTS.md as a standardized format for providing context and instructions to coding agents, with support from major players including OpenAI, Cursor, Google's Jules, and More.

Key details:

Neutral naming: OpenAI Codex team chose AGENTS.md to encourage broad adoption
Industry alignment: Major coding assistants now recognize and use this file
Open source: Microsite and specification available on GitHub
Renaissance moment: Marks shift from fragmented approaches to unified standard
Immediate adoption: Already supported across major tools

I really hope everyone gets on board with a standardized agents file. I've personally built multiple libraries trying to solve this, especially with how different Cursor's agent files were compared to other tools. I think it's a huge win, even if everyone just agrees on a single markdown file. That way we could at least symlink or easily copy to keep multiple agents in sync. I'm definitely working more and more with multiple agents like Claude Code, Cursor, and Gemini all in the same codebase, and I don't even want to think about maintaining rules between them. It's just a painful thing to worry about and manage.

(AGENTS.md Official Site, GitHub Repository)

Browser Use 0.6.0: Breaking Free from Playwright

Browser Use has completely rewritten their web automation framework, moving from Playwright to Chrome DevTools Protocol (CDP) for better performance and stealth.

Technical improvements:

No WebDriver requirement: Direct CDP connection eliminates WebDriver overhead
Less detectable: Removes Playwright fingerprints that sites detect
Faster execution: Direct protocol access improves speed significantly
Better performance: Lower memory usage and faster operations
Maintained compatibility: Existing scripts work with minimal changes

I initially wrote off Browser Use because the Playwright MCP is already so good that I didn't see any reason to use something that leveraged Playwright. But if you look at the Chrome DevTools Protocol and the low level access that gives you to so many things in Chrome, this is definitely a tool I need to revisit for complex scenarios and manipulating the browser. The idea of having access to memory profiling, network interception at the protocol level, and runtime JavaScript compilation control is very enticing, though I haven't had a chance to use these features yet.

(Browser Use GitHub Release)

🛠️ Developer Tooling Updates

OpenAI Codex CLI Works With Your ChatGPT Login

OpenAI's Codex CLI now integrates directly with your ChatGPT Plus subscription, bringing AI-powered coding assistance to your terminal without additional API costs.

This is one of those things that makes your ChatGPT account that much more valuable and brings it closer to parity with Claude Code. As OpenAI's Codex keeps improving, and many people are seeing great results by using it with high thinking turned on, it becomes a much better value proposition to run both Codex and Claude Code in your same workflows.

Claude Code Status Line Customization

The new Status Line feature in Claude Code provides at-a-glance information about your coding session, with various customization options available.

I really love the new Status Line feature of Claude Code. It's a very simple thing, but it helps keep the main information up front in your face. There's lots of ways to customize it, but I just use the default PowerLine one.

(Claude PowerLine GitHub)

💼 AI Ecosystem & Business Updates

Google Expands AI Studio GitHub Integration

Google AI Studio now features direct GitHub integration, allowing developers to commit changes directly from the AI Studio interface.

Integration capabilities:

Direct commits: Push changes without leaving AI Studio
Branch management: Create and switch branches from UI
Authentication: Seamless GitHub account connection
Project sync: Keep AI experiments synchronized with repos
Collaborative workflows: Share AI-generated code with teams

I've loved AI Studio since day one and they're continuing to add amazing features and updates. It's so easy to spike out some idea in the browser and then just copy and paste the generated code and grab a token. It's a beautiful experience and I love seeing it grow. The GitHub integration is going to be really interesting to see how much of a dev tool AI Studio can really turn into. One tip I could give you is that AI Studio is even easy to use on mobile, and it's one of my favorite AI mobile experiences for development.

(Google AI Studio)

OpenAI Launches ChatGPT Go in India

OpenAI has introduced a new affordable tier specifically for the Indian market at Rs. 399 ($4.80 USD), marking their first region-specific pricing strategy.

Subscription details:

Price point: Rs. 399 per month (approximately $4.80 USD)
Features: 10x higher limits than free tier across all metrics
Payment: Native UPI support for Indian users
Localization: All pricing now shown in Indian Rupees
Strategy: Testing ground for global expansion of affordable tiers

ElevenLabs Launches Chat Mode for Voice Agents

ElevenLabs has released Chat Mode, enabling developers to deploy conversational AI agents with natural voice interactions through simple SDKs or a single line of HTML.

Technical capabilities:

Simple deployment: One line of HTML for basic integration
SDK support: Full-featured SDKs for advanced use cases
Music generation: Now includes ElevenLabs Music for audio backgrounds
Real-time processing: Low-latency voice interactions
Cross-platform: Works on web, mobile, and desktop

(ElevenLabs Conversational AI)

💻 Cursor Corner

Cursor CLI Gets Major Update

Cursor CLI now includes MCP support, Review Mode, compression commands, and improved file selection, bringing desktop features to the command line.

New CLI features:

MCP servers: Use .cursor/mcp.json definitions
Review Mode: Ctrl+R for change review and follow-ups
Context compression: /compress to free up context window
@-file selection: Reference files and folders directly
Token counting: Visual indicators for context usage
AGENTS.md support: Automatic recognition of new standard

The race to be the best CLI tool between Claude Code and the Cursor CLI and Codex and Gemini and all of the other options available is almost funny to watch because the one thing AI is really great at is building CLI tools. So I always imagine in my head each company aggressively prompting their AIs to write better CLI tools than their competitors, and borrowing features while at the same time leveraging each other's CLI tools to accomplish these goals. I almost wish they'd all just get together and figure out the best one, but in the end I know competition will bring out the best tools possible.

(Cursor CLI)

New Stealth Model Available in Cursor

Cursor has quietly launched access to a mysterious new model from an unnamed partner, available free to all users for testing.

What we know:

Free access: Available to all Cursor users at no charge
Performance: Early reports suggest strong coding capabilities
Mystery partner: Cursor team keeping the source confidential
Experimental phase: Gathering user feedback for improvements
Integration: Works seamlessly with existing Cursor features

It's fascinating to watch the sheer pace of new models being released. It feels like the "new JavaScript framework every week" era, but now it's a new model every week and it's impossible to keep up. To be honest, the only model updates I'm looking forward to right now are the Anthropic updates and the Google updates. Both are promising significant leaps forward. I've kind of given up trying to follow all the other companies because every time I try them, the tooling just isn't there. No matter how good the model is, if the tooling doesn't support it, then for me it's just not worth it. So I always look to the companies that are invested in both the models and the tooling to make their models perform the best.