AI Agents: Every NEW Claude & Codex Feature Explained
Riley Brown · 2026-05-23
💡 Quick Take
1. Claude now supports terminal interaction with AI agents, allowing multitasking.
2. Andrej Karpathy has joined Anthropic, a significant talent acquisition.
3. Anthropic is making general improvements to its desktop app, aiming for a super app experience.
4. CodeX introduced the `/goal` command for long-running, objective-oriented tasks.
5. CodeX now allows sharing of plugins across teams and workspaces.
6. CodeX has a new "plugin creator" feature to help users build their own plugins.
7. CodeX is rolling out "design mode" features, enabling annotations and direct editing of UI elements.
8. CodeX's "appshots" feature allows users to capture screenshots of any app and immediately send them to CodeX for analysis or action.
9. Google I/O focused heavily on AI agents, but lacked a clear "super app" strategy.
10. Gemini Spark is a promising new feature within Gemini with super app potential, but Google's fragmented approach is a concern.
11. Google's new Gemini 3.5 Flash model is fast but expensive and not considered state-of-the-art.
12. Cursor released a new, fast, and cost-effective model, Composer 2.5, with impressive front-end development capabilities.
13. Cursor is positioning itself as a super app contender for coding and knowledge work, with integrations and automations.
14. The AI agent space is experiencing intense competition, driving rapid platform development and innovation.
15. Chorus.com is a new platform aiming to integrate Claude Code and CodeX into iMessage and other communication channels, functioning as a cloud-based mini super app.
📊 Detailed Explanation
1. Claude now supports terminal interaction with AI agents, allowing multitasking. This is a game-changer for developers and power users who prefer working in the terminal. Instead of a linear chat, you can now fire off multiple agent tasks simultaneously, like researching different topics or checking updates for various tools. The interface allows you to easily navigate between these running tasks, seeing their status and diving into the specific chat for each one. It's about bringing the power of AI agents directly into your command-line workflow, making it more efficient and less disruptive.
2. Andrej Karpathy has joined Anthropic, a significant talent acquisition. This is HUGE news! Karpathy is a rockstar in the AI world, known for his work at Tesla and OpenAI. His move to Anthropic signals a major boost for their research and development efforts, especially in large language models (LLMs). The speculation about the compensation is wild, with some suggesting over a billion dollars, highlighting the immense value placed on top AI talent. This could definitely lead to a migration of talent and potentially influence the competitive landscape, drawing more attention to Claude.
3. Anthropic is making general improvements to its desktop app, aiming for a super app experience. The Claude desktop app is evolving into what the speaker defines as a "super app" – a unified platform for chatting, knowledge work, and coding, with seamless integration to existing tools and the ability to create automations. While the terminal interface offers a unique way to interact, the desktop app aims for broader accessibility and usability. The improvements are making it easier to use "skills" (custom agent functionalities), though the lack of cross-platform skill compatibility (e.g., between Co-work and Claude Code) is still a pain point for some users who want a single, all-encompassing app.
4. CodeX introduced the `/goal` command for long-running, objective-oriented tasks. This is a massive leap for agentic AI. Instead of just executing a command, `/goal` allows you to define an objective for an extended period. The agent stays oriented around achieving that outcome, planning more effectively and working for much longer durations – we're talking hours, even days. It's about moving from task execution to true goal achievement, making AI agents capable of tackling much more complex, multi-stage projects without constant human intervention.
5. CodeX now allows sharing of plugins across teams and workspaces. This is a big deal for collaboration. Developers can now build custom plugins (bundles of apps, skills, and files) and easily share them with their teammates. This streamlines workflows, ensures consistency, and allows teams to leverage specialized AI capabilities built by their own members. The "shared with you" tab makes it easy to discover and use plugins created by others, fostering a more collaborative and efficient AI development environment.
6. CodeX has a new "plugin creator" feature to help users build their own plugins. This democratizes plugin development. You can now ask CodeX to guide you through the process of creating a plugin, even helping to define its skills and interactions with APIs. This lowers the barrier to entry for creating custom AI tools tailored to specific workflows, making the platform even more extensible and personalized.
7. CodeX is rolling out "design mode" features, enabling annotations and direct editing of UI elements. This brings CodeX closer to being a full "vibe coding" platform. While not full visual design like some other tools, it allows users to leave annotations on UI elements (like "make this bigger") which the AI then interprets and applies. You can also directly adjust properties like font size and weight, and the AI will implement these changes. This offers a more intuitive way to interact with and modify visual interfaces directly within the AI environment.
8. CodeX's "appshots" feature allows users to capture screenshots of any app and immediately send them to CodeX for analysis or action. This is incredibly powerful for context-aware AI. By pressing a shortcut (like Command-Command), you can capture a screenshot of whatever app you're using (email, Google Docs, etc.). This screenshot is sent to CodeX, providing immediate context for tasks like drafting responses, researching information, or performing actions within that application. It essentially allows the AI to "see" what you're seeing and act upon it.
9. Google I/O focused heavily on AI agents, but lacked a clear "super app" strategy. While Google talked a lot about AI agents at their big conference, it's unclear what their central platform or "super app" will be. They have multiple initiatives like Anti-gravity, Gemini Spark, and Google AI Studio, but there's no single, cohesive vision for an integrated agentic platform. This fragmentation makes it hard for users to know where to go for specific agentic tasks, unlike the more defined offerings from OpenAI (CodeX) and Anthropic (Claude).
10. Gemini Spark is a promising new feature within Gemini with super app potential, but Google's fragmented approach is a concern. Gemini Spark, with its ability to control folders, spin up virtual computers, connect to apps, and handle agentic tasks, shows a lot of promise as a super app. It has the potential to be a unified platform for various AI-driven tasks. However, Google's continued development of other platforms like Anti-gravity and AI Studio dilutes focus and creates confusion, potentially preventing Spark from reaching its full super app potential compared to more consolidated offerings.
11. Google's new Gemini 3.5 Flash model is fast but expensive and not considered state-of-the-art. While fast, Gemini 3.5 Flash isn't seen as a frontier model, lagging behind competitors like OpenAI's Opus or Anthropic's Claude 3.5. It's noted as being expensive to use and not as capable as some other models, like Cursor's Composer 2.5. This suggests Google is still playing catch-up in the LLM race for cutting-edge agentic capabilities.
12. Cursor released a new, fast, and cost-effective model, Composer 2.5, with impressive front-end development capabilities. This is a big win for Cursor! Composer 2.5 is incredibly fast and cheap to use, making it very attractive for developers. It excels at front-end tasks, generating code quickly and efficiently. The fact that it's significantly faster than CodeX or Claude for certain tasks is a major differentiator.
13. Cursor is positioning itself as a super app contender for coding and knowledge work, with integrations and automations. Cursor is clearly aiming to be a comprehensive AI platform. With its new model, in-app browser, marketplace for integrations, and automations, it's building out the features needed to compete as a super app. While it might not yet have the same document creation capabilities as CodeX or Claude, the stated goal of building the "next generation platform for coding and knowledge work" indicates they are heading in that direction.
14. The AI agent space is experiencing intense competition, driving rapid platform development and innovation. The battle for the enterprise market is fierce, with major players like OpenAI, Anthropic, and Google pouring resources into developing advanced AI agent platforms. This competition is accelerating innovation at an unprecedented pace, leading to rapid improvements in capabilities, features, and user experience across the board. The result is increasingly powerful and versatile AI tools becoming available to users.
15. Chorus.com is a new platform aiming to integrate Claude Code and CodeX into iMessage and other communication channels, functioning as a cloud-based mini super app. This is a really cool initiative to bring AI agents into everyday communication. Chorus.com allows you to connect your AI subscriptions (like ChatGPT) to create personalized agents that can operate within iMessage, and soon WhatsApp and Slack. It functions as a mobile-first, cloud-based mini super app, enabling tasks like building landing pages directly from your phone via text. It offers flexibility in choosing underlying models and aims for a proactive, automated experience.
🎯 Expert Opinion
Wow, what a whirlwind of AI agent updates! It's absolutely electrifying to see the pace of innovation right now. The core theme emerging is the race towards the "super app" – a single, unified platform where users can seamlessly handle a multitude of tasks, from coding and knowledge work to complex long-running projects. OpenAI's CodeX is making serious strides here with features like `/goal` and plugin sharing, really solidifying its position as a powerful, collaborative environment. The introduction of `/goal` is particularly significant; it signals a shift from AI as a task executor to AI as a true goal-achiever, which is crucial for tackling real-world complexity.
Anthropic's acquisition of Andrej Karpathy is a masterstroke. This isn't just about adding a brilliant mind; it's a statement of intent. Karpathy's deep understanding of LLMs and his passion for education will undoubtedly propel Anthropic's R&D forward, potentially creating a formidable competitor to OpenAI. The speculation around his compensation underscores the immense value of top-tier AI talent and the fierce competition to secure it. This move could very well shift the balance of power in the LLM landscape.
Google's situation, however, is a bit of a mixed bag. They are clearly investing heavily in AI, as evidenced by the Google I/O announcements and Gemini Spark. Gemini Spark, in particular, has the *potential* to be a game-changer, offering a centralized hub for agentic tasks. But the fragmentation across Anti-gravity, AI Studio, and Gemini itself is a major concern. As an expert, I see this as a classic case of a large organization struggling to consolidate its efforts. Unless Google can coalesce around a single, dominant super app strategy, they risk diluting their impact and falling behind more focused competitors like OpenAI and Anthropic. Their current models, like Gemini 3.5 Flash, while fast, aren't pushing the frontier, which is a worrying sign given their resources.
Cursor's emergence is also fascinating. The acquisition by SpaceX provides them with the computational power to develop truly cutting-edge models like Composer 2.5. This model's speed and cost-effectiveness, especially for front-end development, make it a compelling alternative. Cursor's ambition to become a super app for coding and knowledge work, complete with integrations and automations, positions them as a serious contender. I expect them to rapidly gain market share, especially among developers looking for optimized workflows.
The overarching trend is clear: the AI agent space is becoming a battleground. Companies are pouring resources into building comprehensive platforms that integrate deeply into user workflows. This intense competition is fantastic for us as users, as it means we'll see faster innovation, more powerful features, and increasingly sophisticated AI capabilities. The concept of a "super app" is no longer a futuristic idea; it's the immediate future, and the companies that can deliver a truly seamless, integrated experience will dominate. I'll be watching Chorus.com with keen interest too – bringing these powerful agents into familiar communication channels like iMessage is a smart move that could unlock new levels of accessibility and utility.
Kanal: Riley Brown