Welcome back. The team at Cursor has been cooking. The company just unveiled Composer 1 — their first coding model and what they called "the best way to code with agents." Composer 1 is cheaper than Claude, can design beautiful frontends (like Claude, unlike Codex), and most importantly, it's blazingly fast. You can learn more about Composer and Cursor 2.0 in this 1-hour tutorial.

Welcome to The Code. This is a 2x weekly email that cuts through the noise to help devs, engineers, and technical leaders find high-signal news, releases, and resources in 5 minutes or less. You can sign up or share this email here.

Cursor launches its own coding model that's 4x faster: Composer is the startup’s first proprietary model, and it comes with a redesigned interface called Cursor 2.0. Composer wraps up most coding tasks in under 30 seconds — four times faster than similar models — and was trained to better understand large codebases. The new interface lets you run multiple AI agents at the same time without them getting in each other's way. Here’s a 1-hour tutorial on how to use it.

OpenAI's new safety models let developers write their own rules: With gpt-oss-safeguard, OpenAI is giving developers unprecedented control over content moderation. The two open-weight models (120B and 20B) can classify text against any developer-provided policy — whether you're running a gaming forum that needs to catch cheating discussions or a review site filtering fake posts. Here’s OpenAI’s official guide on how to use it.

Cognition launches a new model optimized for software engineering: The team behind Devin just dropped SWE-1.5, a powerful AI model that hits near-top performance on coding tasks while running at up to 950 tokens per second. That's 6x faster than Haiku 4.5 and 13x faster than Sonnet 4.5. It's already live in Windsurf, where Cognition's own engineers now use it daily for digging through codebases and building apps.

Google Cloud CTO's playbook for scaling AI agents: Building agents isn't magic— it requires documented business processes and incremental measurement, says Google Cloud CTO Will Grannis. His team breaks complex agent workflows into "very atomic, very specific" steps with measurable outcomes at each gate, deciding whether to proceed, pivot, or stop.

Engineering manager’s exact 1:1 script: Most engineering managers wing their one-on-ones, but Can Duruk has a better approach. His weekly structure uses the People/Product/Process framework to cover team dynamics, work progress, and systemic issues.

Prioritization Starts With Strategic Prioritization: Product consultant John Cutler's latest post tackles a common problem — teams being asked to prioritize without clear strategy. He shows how every decision should tie back to four business fundamentals while fighting natural market decline.

Cursor AI agents work like 10 developers (demo): A VP from Cursor just walked through his daily coding workflow in a hands-on tutorial. The secret: break work into bite-sized tasks and let AI agents handle different jobs — one agent fixes bugs, another handles security checks.

OpenAI releases guide for building custom code reviews with Codex SDK: Developers working with on-prem repositories or non-GitHub platforms can now build their own automated code review systems using OpenAI's Codex CLI. The new guide walks teams through installing Codex in CI/CD runners, using headless mode with structured JSON outputs, and integrating review comments directly into pull requests.

Fine-tuning and Reinforcement Learning for LLMs: Intro to Post-TrainingThis new course, taught by Sharon Zhou, VP of AI at AMD, teaches how to turn pre-trained LLMs into reliable, production-ready systems. In 5 modules, you’ll walk through the full post-training pipeline with practical techniques used by top labs.

Signs of introspection in LLMs: This research investigates whether large language models like Claude can genuinely introspect on their internal states or merely fabricate plausible responses when queried. Key finding: Claude shows limited but authentic introspection, accurately detecting and describing injected "thoughts" in experiments, though performance varies with injection strength.

Agent Data Protocol: This research introduces the Agent Data Protocol (ADP), a unified format to combine 13 diverse agent datasets into the largest SFT collection (1.27M trajectories) for training agentic language models across coding, browsing, and tool use. ADP enables ~20% average performance gains, achieving SOTA or near-SOTA results without domain-specific tuning.

Fundamentals of Building Autonomous LLM Agents: This paper reviews the architecture and methods for building autonomous agents using LLMs. Key finding: Integrating perception, reasoning, memory, and execution systems enables LLMs to mimic human cognition, creating more adaptive and capable agents that bridge AI-human performance gaps.

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

Until next time — The Code team