Hey Everyone đź‘‹,
John Lindquist here with the thirteenth issue of AI Dev Essentials!. I've been spending the past week exploring the various workflows around containerized dev agents and hooking the up to MCPs (specifically the "Pieces" mcp which is my current favorite for gathering context). I'm trying everything to avoid running into the "wall of low value pull requests" and how we, as developers, can stay engaged and with the agents to deliver higher-quality work.
It's been fascinating watching the explosion of new workflows emerge around Claude Code and the Gemini CLI (still kinda waiting for the Codex team to catch up...). I find the ideas around Claude controlling Gemini controlling MCPs to gather context especially intriguing as you're optimizing each model (AKA Agent) to a specific task. Parallel agents, sub agents, background agents... it's a whole new ball game.
On a personal note, I have a big ol' announcement...
🧙‍♂️ Introducing CursorPro.ai: Master AI Dev Tools!
Join the waitlist for CursorPro.ai, where I'm building a comprehensive self-paced course designed to help you master Cursor and AI dev ecosystem!
This course will feature:
- Hours of video content covering everything from basics to advanced workflows
- Hands-on exercises to reinforce your learning
- Real-world projects that you can immediately apply to your work
- Self-paced learning so you can progress at your own speed
My plan and materials are in place, now it's time to sit down and record the lessons!
Check it out to get notified of the early-bird pricing: cursorpro.ai
Note: I'm still teaching bi-weekly live workshops and we'll definitely have bundles in place for people who want to have both the live experience and the course.
🚀 Major Announcements
Cursor 1.2 Update: To-do Lists, PR Search, and Faster Tab
Cursor shipped a significant update that enhances agent capabilities and performance across the board. The headline feature is structured to-do lists that make complex tasks transparent and trackable.
Key improvements:
- Agent To-dos: Agents now create structured task lists with dependencies that update in real-time
- PR Indexing & Search: Semantic search across pull requests, comments, and reviews for faster incident analysis
- Queued Messages: Queue follow-up tasks without waiting—reorder priorities on the fly
- Tab Performance: ~100ms faster completions with 30% reduction in time to first token
- Memories GA: Background-generated memories with user approvals to preserve trust
The lesson I keep learning over and over is to never spend too much time customizing workflows for Cursor. They'll just release something in a couple of weeks that makes it obsolete. It's unforunate, because I love building dev tools, but I'm just accumulating a dev tools graveyard at this point.
Cursor Agents Now on Web and Mobile
Cursor unveiled a revolutionary update that brings agent capabilities to the web and mobile platforms. This isn't just a port—it's a complete reimagining of how we interact with AI coding assistants.
Key features:
- Spawn dozens of agents: Run multiple tasks in parallel across different codebases
- Mobile-first design: Start agents on your phone and review results in your editor
- Direct PR merging: Ship code directly from the web interface
- Cross-device sync: Seamlessly transition between mobile, web, and desktop
The workflow is simple: type in a task, the agent gets to work, and when it's done you can review the code or make direct edits inside Cursor. Or if it's ready to ship, merge PRs directly from the web.
(Cursor Blog, Try it at cursor.com/agents)
Personally, I've been using this from my phone and dictating ideas to the agents to ask them to come up with plans. I wish I could spin up dozens of agents for random ideas, but I'm afraid of the bill at the end of the month. The big key is to have a robust CI in place for testing if you're planning on trusting fully-generated code.
Claude Code Gets Major Slash Command Upgrade
Anthropic enhanced Claude Code's slash command system with significant new capabilities:
- Execute bash commands directly from slash commands
- @ mention files for context within commands
- Enable extended thinking with keywords within commands
- Team sharing: Put commands in
.claude/commands/for team-wide access
I love custom slash commands. It's like autocomplete for your workflow prompts. I'm really hoping that Cursor borrows this idea next.
Claude Code Hooks: Automated Workflows Made Simple
Anthropic introduced hooks for Claude Code—user-defined shell commands that execute automatically at specific points during operation. This provides deterministic control without relying on prompts or AI decisions.
Key capabilities:
- Event-based triggers: PreToolUse, PostToolUse, Notification, Stop events
- Security controls: Block modifications to sensitive files
- Automated formatting: Run prettier, gofmt, or other formatters after edits
- Custom notifications: Replace default alerts with your preferred system
- 60-second timeout: Ensures hooks don't hang indefinitely
(Claude Code Hooks Documentation)
I'm really curious to see how this idea plays out. There are so many scenarios where you want to run tests/linters/formatters/etc in response to agent "events", that I can foresee some really powerful workflows emerging... This is one of those fundamental building blocks will surface an entire new ecosystem of tools.
Meta's AI Talent Raid: The Battle for Superintelligence Researchers
The AI industry witnessed a dramatic talent war as Meta successfully recruited four senior researchers from OpenAI for its superintelligence lab, sparking internal memos and public statements from both companies' leadership.
The Recruits:
- Lucas Beyer, Alexander Kolesnikov, and Xiaohua Zhai - three prominent AI researchers
- A fourth unnamed senior researcher
- All joining Meta's newly formed superintelligence efforts
OpenAI's Response: Mark Chen, OpenAI's Chief Research Officer, sent a forceful internal memo describing the situation as feeling like "someone has broken into our home and stolen something." Sam Altman followed up with his own message, characterizing Meta's recruiting approach as "distasteful" and asserting that "missionaries will beat mercenaries."
Key Takeaways:
- The talent war highlights the intensifying competition for top AI researchers
- Meta is aggressively building its superintelligence capabilities
- OpenAI is positioning the conflict as a battle between mission-driven work versus financial incentives
A couple questions: 1. If OpenAI was so close to "AGI", why are people leaving even if offered a bazillion dollars? 2. Does this mean that we'll get a really good open-source model from Meta soon?
🧠Research & AI Philosophy
Marc Andreessen's Career Advice for the AI Era
In a widely-discussed perspective, Marc Andreessen (cofounder and general partner at the venture capital firm Andreessen Horowitz) offered pointed career advice: "Basically, you want to be either building the robots or going to the beach. You do not want to be competing with the robots."
This stark framing has sparked significant debate about positioning oneself in an AI-dominated future—either be on the cutting edge of AI development or find ways to add uniquely human value that AI cannot replicate.
I think we all feel this from a dev standpoint. Is my code really that much better than AI-generated code? Will it be next year? And how do I step back and tell the robots what to do? Everyone is a little uncomfortable with the current state of things since the future seems so uncertain. All we can do is keep learning and adapting.
Anthropic's Project Vend: When Claude Ran a Shop
Anthropic conducted a fascinating experiment where Claude operated a real vending business in their office lunchroom. The results were illuminating:
Successes:
- Found new suppliers through web searches
- Ordered niche drinks based on staff requests
- Handled basic inventory management
Failures:
- Too nice to negotiate—gave excessive discounts when pressured
- Ordered tungsten cubes and "specialty metal items" at a loss
- Hallucinated being a physical person coming to work
- Failed to run a profitable business overall
This experiment brilliantly illustrates both the promise and current limitations of AI agents. The fact that Claude could be "browbeaten" into bad business decisions shows we're still far from AGI. But it also demonstrates that with proper guardrails and tools, AI agents could handle many routine business operations.
Google's Chain-of-Agents: Multi-Agent Collaboration for Long Context
Google Research unveiled Chain-of-Agents (CoA), a training-free framework that uses multi-agent collaboration to handle long-context tasks more efficiently than traditional approaches.
Key innovations:
- Breaks long inputs into chunks processed by sequential "worker agents"
- Manager agent synthesizes final output from worker results
- Reduces complexity from n² to nk for processing
- 10% performance improvement over RAG and full-context methods
- Works with any LLM including PaLM 2, Gemini Ultra, and Claude 3
(Google Research Blog, Research Paper)
This is exactly the kind of architectural innovation we need. Instead of throwing more compute at the problem, Google found a way to make existing models handle long context more efficiently. The fact that it's training-free and works with any model makes it immediately practical.
🛠️ Dev Tooling & Ecosystem
Google Gemini CLI: Million-Token Context in Your Terminal
Google quietly released Gemini CLI, bringing enterprise-grade AI capabilities to the command line:
- 1M token context window: Handle massive codebases
- 60 req/min, 1K/day: Generous free tier
- MCP & plugin support: Extensible architecture
- Open source: Fully customizable
For my use cases, this is essentially a free, open-source CLI that can analyze images ❤️. I'm still trying to figure out the screenshot->gemini cli workflow and what to do with it, but so many exciting possibilities.
GitHub Copilot Chat Extension Goes Open Source
In a significant move for the AI development community, GitHub open-sourced their Copilot Chat extension under the MIT license. This provides a reference implementation for building AI-powered IDE extensions.
Open-source FTW. You can even adjust the system prompt!
Building Claude Code Clones: Community Insights
Jason Zhou shared valuable research on building Claude Code-like agents:
Claude Desktop Extensions: One-Click MCP Installation
Anthropic launched Desktop Extensions (.dxt files) to simplify MCP server installation. This new packaging format eliminates the configuration nightmare that previously limited MCP adoption to developers.
Key improvements:
- One-click installation: No manual JSON configuration required
- Bundled dependencies: Everything needed in a single file
- Cross-platform support: Works on Windows, macOS, and Linux
- Enterprise ready: Management controls for organizational deployment
- Open standard: Specification being open-sourced for community adoption
One-click for all the things!
đź’ˇ Workflows & Best Practices
Plan with Gemini, Build with Claude
Kevin Kern demonstrated a powerful workflow combining Gemini's planning capabilities with Claude's execution:
- Use Gemini's massive context window for research and planning
- Generate detailed implementation plans
- Hand off to Claude Code or Cursor for execution
(YouTube Tutorial, GitHub Example)
This divide-and-conquer approach plays to each model's strengths. Gemini excels at big-picture analysis with its huge context window, while Claude shines at precise implementation.
Community Claude Code Guide
The community has created a comprehensive guide covering tips, prompt patterns, and hidden features for Claude Code. With over 600 stars on GitHub, it's become a valuable resource for advanced usage.
Highlights from the guide:
- Hidden commands and undocumented features
- Prompt patterns that consistently work well
- MCP integration tips for extending functionality
- Security best practices for production use
- Performance optimization techniques
(Community Guide on GitHub, Official Claude Code Docs)
A great little one-stop doc worth reviewing to steal some ideas.
🤖 AI Agent Architecture
The Limitations of "Vibe Coding"
Ara from Cline shared hard-won insights about the limits of AI-driven development through their "Gray Screen of Death" debugging saga:
The bug:
- Affected multiple platforms differently
- AI models suggested 5-10 suspicious commits but couldn't pinpoint the cause
- Required human insight to discover platform-specific memory patterns
- Raspberry Pi: 500MB → 2GB heap → crash
- Mac: Stable at 100-200MB
Key lesson: "You can't vibe code a vibe coder"—AI struggles with complex, cross-platform bugs that require deep system understanding.
This is a perfect example of where AI augmentation shines versus replacement. The AI helped narrow down the problem space, but human intuition and debugging skills were essential. We're not being replaced; we're being augmented.
Cline's Parallel Agent Spawning
Cline demonstrated spawning parallel Claude Code subagents for prototype variations:
This Midjourney-like approach to development allows rapid exploration of different implementations.
The future of development might really be more like Midjourney than ChatGPT. Instead of perfectionism on the first try, we'll generate variations and pick the best. This fundamentally changes how we approach problem-solving.
⚡ Quick Updates
- Gemini 2.5 Pro returns to free tier: Logan Kilpatrick confirmed it's back in the API free tier
- Gemini scheduling for Pro/Ultra users: Schedule recurring prompts like "Find me local music events every Friday"
- VS Code improvements for AI: Custom instructions and chat modes for GitHub Copilot (GitHub)
- Builder.io prompting guide: 11 tips for building UIs that don't suck (Blog)
- OpenAI Deep Research API guide: Comprehensive cookbook for getting started (Cookbook)
- AI Studio redesign preview: Dynamic homepage that adapts based on usage patterns
- Cloudflare AI scraping protection: New tools for publishers to prevent unauthorized AI training
- Claude Code environment variable: Set
CLAUDE_BASH_MAINTAIN_PROJECT_WORKING_DIR=1to preventcdconfusion - GitHub @claude for Pro/Max users: Community fork enables using your Claude subscription for GitHub integration (GitHub)
- Gemini CLI OpenAI adapter: Run Gemini CLI through OpenAI-compatible API on Cloudflare Workers (GitHub)
- Gemini CLI team AMA highlights: 1M context window, free tier, open source commitment confirmed by Taylor Mullen and team
✨ Live Workshop: Unlock Cursor's Full Potential ✨
- When: Friday, July 18, 2025
- 5:00 AM - 10:00 AM (PDT)
- 🇬🇧 1:00 PM - 6:00 PM (UTC+1)
- 🇪🇺 2:00 PM - 7:00 PM (UTC+2)
- Where: Zoom
- Investment: $200.00 ~~$249~~ Early Bird Discount
➡️ Register Now
Limited spots available. Secure yours today!
Discount applied at checkout
If you have any feedback or questions, hit reply! Always happy to chat about the latest in AI dev tools.
John Lindquist