<p><strong>TLDR:</strong> Cursor’s Composer 2 was exposed as a fine-tuned Kimi K2.5 model after a developer spotted the model ID <code>kimi-k2p5-rl-0317-s515-fast</code> in API calls. Moonshot AI confirmed it, and the community reaction ranged from “smart move” to “why hide it.” Meanwhile, OpenAI is consolidating ChatGPT, Codex, and its internal browser into a single desktop app aimed squarely at enterprise. Claude Code picked up scheduled cloud tasks and DOM element selection on desktop today.</p>
<hr>
<h2>Cursor’s Composer 2 is Kimi K2.5 with RL on top</h2>
<p>Someone sniffed Cursor’s API calls and found the model ID <code>kimi-k2p5-rl-0317-s515-fast</code> sitting under Anysphere’s account. Moonshot AI confirmed it: Composer 2 is Kimi K2.5 with reinforcement learning, not a proprietary in-house model. The controversy isn’t really about using an open-source base — that’s a reasonable engineering call, and Kimi K2.5 scored well on perplexity-based evals. The problem is Cursor was raising at a reported $50B valuation partly on the claim of in-house model development, and they launched Composer 2 without mentioning any of this. Hugging Face’s Clement Delangue called it a win for open source. Others pointed out Cursor didn’t file the serial numbers off — the model ID still says Kimi.</p>
<h2>OpenAI consolidating ChatGPT, Codex, and Atlas into one desktop app</h2>
<p>OpenAI is building a desktop app that merges ChatGPT, its Codex coding platform, and an internal AI browser called Atlas. Greg Brockman is temporarily overseeing the overhaul. The strategic logic is clear: Anthropic is reportedly winning around 70% of head-to-head enterprise deals, and a fragmented product surface isn’t helping. The tension worth watching is whether OpenAI can serve both the consumer chatbot audience that made ChatGPT famous and the enterprise/developer buyers they’re now chasing — those are genuinely different products with different requirements. Enterprise governance features like identity management and audit trails for agents aren’t something OpenAI has addressed publicly in detail yet, and that’s table stakes for the deals they want.</p>
<h2>Claude Code adds scheduled cloud tasks and DOM selection</h2>
<p>Two useful Claude Code updates today. Scheduled cloud tasks let you set a repo, a schedule, and a prompt — Claude runs it automatically without you keeping a browser tab open. Separately, the desktop client now lets you click a DOM element directly instead of describing it in prose. For React apps it also pulls the source file, component name, and props. Small change, but anyone who’s typed “the blue button in the top-right corner of the nav” more than twice will appreciate it.</p>
<h2>Karpathy: no manual code since December</h2>
<p>Andrej Karpathy said on a recent podcast that he hasn’t typed a line of code since December, delegating almost entirely to agents. He described running multiple Claude Code and Codex windows in parallel, discussed AutoResearch loops that outperformed his own two decades of hyperparameter tuning, and reflected on LLM “jaggedness” — models excelling at verifiable tasks while stagnating on unverifiable ones. Worth reading the full thread for his take on model speciation and distributed compute pools for collaborative AI research.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Spotify’s Honk agent</strong> is merging 1,000 pull requests every 10 days for large-scale code migrations across their codebase.</li>
<li><strong>Stripe’s Minions</strong> generate over 1,300 pull requests per week from single instructions, up from 1,000 in earlier trials, all human-reviewed with no human-written code.</li>
<li><strong>Cognition (Devin)</strong> reports 70% of sessions are currently human-initiated, predicting that flips to 30/70 within months and 10/90 within a year.</li>
<li><strong>Next.js 16.2</strong> ships with <a href="http://AGENTS.md" target="_blank" rel="noopener nofollow">AGENTS.md</a> and bundled docs so agents have deep knowledge of the exact framework version in use, plus a <code>@vercel/next-browser</code> package for frontend debugging.</li>
<li><strong>Google Stitch</strong> now generates full app UI designs from a single text prompt, with MCP export for Cursor and Claude Code and Figma integration.</li>
<li><strong>Windsurf pricing change</strong> drew customer complaints after prompt credits were replaced by daily quotas without notice or consent.</li>
<li><strong>Amazon’s AI outage fallout</strong> has resulted in a policy requiring senior engineer (L5+) sign-off on all code changes from junior and mid-level developers — creating a review bottleneck that may offset the speed gains from AI tooling.</li>
<li><strong>GitHub Copilot</strong> now resolves the generic “Auto” label to actual model names in usage metrics, and agent commits include an <code>Agent-Logs-Url</code> trailer for tracing why specific changes were made.</li>
<li><strong>Grok 4.20</strong> scored 6.0% on CritPt, roughly 2x better than DeepSeek V3.2 and close to Speciale, according to one benchmark thread.</li>
<li><strong>Sonatype Guide</strong> launched as an MCP server that intercepts AI-generated dependency suggestions and flags hallucinated, vulnerable, or outdated packages before they land in your code.</li>
<li><strong>OpenCode</strong> now supports Composer 2 via Cursor subscriptions, added shift-enter via the kitty keyboard protocol, and picked up a Firecrawl plugin for web scraping from agents.</li>
</ul>


AI coding assistants, code generation tools, and vibe coding workflows.

Agentic Digest

Cursor's Composer 2 was revealed to be a fine-tuned Kimi K2.5 model after a developer spotted the model ID in API calls, raising questions about transparency given Cursor's reported $50B valuation and claims of in-house model development. OpenAI is merging ChatGPT, Codex, and its internal Atlas browser into a single desktop app as it tries to compete with Anthropic in enterprise deals. Claude Code added scheduled cloud tasks and DOM element selection for desktop. Andrej Karpathy shared he hasn't written manual code since December, relying entirely on agents. Notable industry updates include Spotify's Honk agent merging 1,000 PRs every 10 days, Stripe's Minions generating 1,300+ PRs per week, Next.js 16.2 shipping with AGENTS.md support, and Windsurf drawing complaints over an unannounced pricing change.

Cursor's Kimi K2.5 cover blown, OpenAI merges ChatGPT and Codex into one app