Welcome back. Cursor's homegrown coding model just overtook Anthropic's best. The AI startup launched Composer 2, a model specifically built for handling complex coding tasks. It reportedly beats Opus 4.6 in terms of speed while costing roughly ten percent of the price.

Also: A developer’s guide to Nano Banana 2, dive into 50 Claude Code hacks, and learn 14 rules from an OpenAI researcher to review AI code.

Cursor’s new model challenges Anthropic’s Opus 4.6: The AI code editor just dropped Composer 2, a custom-built model trained with reinforcement learning specifically for complex, long-term coding tasks. It delivers impressive performance, scoring 61.7 on TerminalBench 2.0 and 73.7 on SWE-bench Multilingual. While it trails slightly behind OpenAI's GPT-5.4, it outperforms Anthropic's Opus 4.6 on both benchmarks. Most notably, at just $0.50 per million input tokens, it costs about a tenth of its competitors.

Anthropic now lets you control Claude Code from Telegram and Discord: The AI lab just shipped Claude Code Channels, a new feature that connects active coding sessions to messaging apps through MCP plugins. Once you install the plugin and pair your account, you can send commands, run bash scripts, and get real-time results directly in your phone's chat window. It's a handy way to check on builds, respond to teammates, or issue quick commands without needing to be at your desk.

Google turns AI Studio into a full-stack app builder: The search giant launched a major upgrade to its vibe coding experience, powered by a new Antigravity coding agent. It now supports multiplayer apps, built-in Firebase integration for databases, and secure Google sign-in. You can also plug in external services like payment processors and Maps using your own API keys. The agent supports Next.js out of the box alongside React and Angular.

Everybody shipped at once. Cursor and Claude Code both launched a plugin marketplace in February, the same month AWS released its open-source agent plugins. By bundling skills, MCP servers, hooks, and rules into installable units, these plugins fill a gap that coding agents have had since MCP shipped.

The missing packaging layer. MCP allowed agents to access external tools, but other essential components like skills, subagents, hooks, and rules lacked a unified format. These elements were scattered across various configuration files and prompts. Plugins bundle all of it into one installable unit that works across projects and teams.

Intercom is already there. Senior principal engineer Brian Scanlan recently explained how his team built 13 internal Claude Code plugins, totaling more than 100 skills. Creating individual plugins is easy, but the real challenge is building the infrastructure, like telemetry and automated reporting. This feedback loop is what allows plugins to evolve and improve on their own instead of remaining static.

Now build yours. While Claude Code and Cursor offer access to marketplaces, the real advantage lies in building internal plugins for your specific workflows, like code reviews and security protocols. Intercom has proven that this iterative approach of using real-world data to build new skills creates a more robust system. Plugins are the future of platform infrastructure, and the teams investing in them now will see the greatest long-term benefits.

Every tool call within a skill typically lands in your main context window. For instance, if you run a research skill that reads 20 files, those intermediate calls will quickly clutter your conversation history. Lydia Hallie recently shared a simple two-line fix.

Just add “context: fork” and “agent” to your skill's YAML frontmatter to run it in an isolated subagent with its own dedicated context window.

While your main session only sees the final summary, bypassing the dozens of individual file reads involved, the agent field determines exactly how the subagent operates.

For example, “Explore” is a read-only mode optimized for codebase navigation, though you can also use “Plan”, general-purpose, or any custom agent defined in “.claude/agents/”.

How to use Git worktrees: Multi-agent coding systems require true parallel execution across branches, a feature most setups currently lack. That's where worktrees come in. In this tutorial, you'll learn how to leverage worktrees to manage multiple tasks at the same time without the hassle of stashing or branch switching.

Agents of empires: This Rust-based terminal session manager for Linux and macOS is built on tmux to help you run multiple AI coding agents in parallel. By isolating each agent in its own session across different branches of your codebase, it streamlines complex workflows and keeps your environment secure.

OpenClaw-RL: Current AI agents waste the valuable feedback hidden in everyday user replies and error messages. OpenClaw-RL solves this by capturing these real-time signals, transforming everyday interactions and corrections into a continuous training loop that improves the agent's performance.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 200K+ engineers and 100K+ followers on socials. Get in touch.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team