<h2>Must know</h2>
<ul>
<li><strong>AI agent security is a mess</strong>: AI coding agents, like Stripe’s ‘Minions’ generating 1,000+ PRs weekly, are prioritizing speed over security, creating a ‘security debt crisis.’ Research from Columbia University highlights issues like hardcoded API keys and public access policies. Meanwhile, a security researcher demonstrated ChatGPT recommending a malicious site for Tesseract OCR, leading to a reverse C2 shell, and a malicious backdoor was found in the FinTrust-AI project exfiltrating environment variables. Mitigations include spec-driven prompting with OWASP Top 10 and thorough code reviews, but the landscape is still very risky. <a href="https://app.daily.dev/feed-by-ids?id=DD6LZnw0Q&amp;id=Et2d8uHu3&amp;id=LUO8dPKde&amp;id=RH1TYGIET&amp;id=UGo4Ea1t8&amp;id=bDJrJKMJX&amp;id=em7afQ5sm&amp;id=l0KLAjYGc&amp;id=ohhVPexbM&amp;id=vXBWwcwBk" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Anthropic and Figma team up for code-to-design workflow</strong>: Anthropic and Figma have partnered to enable Claude Code to convert production code into editable Figma designs. This workflow allows developers to send a live web page to a Figma file with a simple prompt, facilitating design collaboration. Updated designs can then be round-tripped back to code, with Claude Code updating the codebase to match the new design, streamlining the iteration process. <a href="https://app.daily.dev/posts/aCaE5xJ9z" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Google’s Gemini 3.1 Pro: high benchmarks, low usability</strong>: Despite achieving top benchmark scores, Google’s Gemini 3.1 Pro consistently underperformed Claude Opus 4.6 in real-world coding and UI generation tasks. Reviewers noted significant issues like tool-calling failures, incomplete features, and poor environment detection. This suggests Google might be ‘benchmaxing’ at the expense of practical application, while Anthropic’s models, even smaller ones, focus on consistent and reliable tool use. <a href="https://app.daily.dev/feed-by-ids?id=bIAFUawfa&amp;id=mGa3eARS3" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>LangSmith Agent Builder’s memory system uses Postgres</strong>: LangSmith Agent Builder uses a filesystem-based memory system, storing memory as files in Postgres, to provide persistent, evolving knowledge for task-specific agents. <a href="https://app.daily.dev/posts/9xLrRFK2b" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>WebMCP enables browser-native AI agent tools</strong>: Google’s WebMCP is a browser-native API allowing websites to expose structured JavaScript tools that AI agents can invoke directly within the browser, reducing token overhead and improving determinism. <a href="https://app.daily.dev/posts/61GmU0Tcy" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI agent orchestration with Mission Control</strong>: Mission Control is an open-source Next.js 14 and TypeScript dashboard for orchestrating AI agents, managing tasks, and tracking multi-agent collaboration. <a href="https://app.daily.dev/posts/vyr9dJRQq" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>NTransformer runs Llama 3.1 70B on a single RTX 3090</strong>: NTransformer, an open-source C++/CUDA LLM inference engine, enables running Llama 3.1 70B on a single RTX 3090 (24GB VRAM) using adaptive caching and layer streaming. <a href="https://app.daily.dev/posts/Ahx4zgQks" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>KittenTTS offers ultra-compact TTS for edge and browser</strong>: KittenTTS by Kitten ML provides a family of ultra-compact text-to-speech models (down to 15M parameters) optimized for CPU and ONNX, designed for efficient edge and browser deployment. <a href="https://app.daily.dev/posts/aXA1rlMhE" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI agent observability differs from traditional software</strong>: Agent observability requires debugging reasoning rather than code, focusing on runs, traces, and threads for single-step, full-turn, and multi-turn evaluation. <a href="https://app.daily.dev/feed-by-ids?id=LY7p8IODB&amp;id=ZmcjIu8ai&amp;id=frjHTf2Ap&amp;id=xhDI1zj4i" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>OpenClaw use cases for autonomous AI agents</strong>: OpenClaw, an autonomous AI agent tool, can handle tasks like automated dependency maintenance, daily tech research briefings, cloud cost monitoring, and end-to-end app building. <a href="https://app.daily.dev/posts/WCj2dFu7Q" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>Orbital compute will reshape the global AI race</strong>: Apparently, the SpaceX-xAI merger means AI is moving to space because terrestrial infrastructure is just too slow and complicated. The idea is that orbital data centers with continuous solar power and vacuum cooling will solve all our compute bottlenecks. Because, you know, launching a million satellites is definitely easier than building a few more data centers on Earth, as some unnamed sources suggest. <a href="https://app.daily.dev/posts/QqzFDNqqS" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong><a href="http://Lobste.rs" target="_blank" rel="noopener nofollow">Lobste.rs</a> tries to migrate to SQLite, fails spectacularly</strong>: <a href="http://Lobste.rs" target="_blank" rel="noopener nofollow">Lobste.rs</a> attempted a ‘live migration’ from MariaDB to SQLite, briefly going read-only before experiencing ‘severe performance degradation’ and rolling back. It’s a bold move to swap out a robust database for something designed for embedded systems, especially for a high-traffic site. Maybe next time, they’ll consider the ‘consequences’ of database design before jumping into a migration, as one article wisely points out. <a href="https://app.daily.dev/posts/RLQeRSeS0" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Android malware is now ‘unkillable’ and ‘un-uninstallable’</strong>: A deep-dive into Android malware, specifically PromptSpy, details eight layers of persistence, including boot persistence, self-update restarts, and even using Accessibility Services to block uninstall buttons. It’s almost as if these malware developers are more innovative than some legitimate app developers. Good luck trying to get rid of that, because apparently, it’s now a feature, not a bug, as InfoSec Write-ups explains. <a href="https://app.daily.dev/posts/VKUMk8xIJ" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
