<h2>Must know</h2>
<ul>
<li><strong>AI agents are still struggling with real-world tasks</strong>: A new benchmark, Apex-Agents, from Mercor, reveals that even leading AI models like Gemini 3 Flash and GPT-5.2 achieve only 24% accuracy on complex white-collar tasks involving multi-domain reasoning and common workplace tools. This suggests that while AI agents show promise in controlled demos, they are far from reliably replacing knowledge workers in real-world scenarios, highlighting a significant gap between current capabilities and practical application. Anthropic’s own engineering team is struggling to design technical evaluations that their Claude models can’t ace, forcing them to create unconventional puzzles to find tasks where human reasoning still holds an edge, as reported by Anthropic. <a href="https://app.daily.dev/feed-by-ids?id=OsmrZ9UpM&amp;id=vA7gH0M1s" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>eBay bans unauthorized AI shopping agents</strong>: eBay has updated its User Agreement to explicitly prohibit third-party AI agents and chatbots from making purchases without permission, effective February 20, 2026. This move directly addresses the rise of autonomous AI tools, like ChatGPT’s Instant Checkout, that can browse and buy products without human review. The policy reflects growing concerns about unchecked automation and potential risks in e-commerce, signaling a pushback against AI agents operating freely on major platforms. <a href="https://app.daily.dev/posts/eNz11CwvB" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Google AI mode reads your Gmail and photos</strong>: Google has launched ‘AI Mode with Personal Intelligence’ for AI Pro and AI Ultra subscribers in the U.S., allowing the AI to access users’ Gmail and Google Photos if they opt-in. This feature can generate vacation itineraries from hotel bookings in emails or suggest shopping items based on travel plans. Google states the AI only processes specific prompts and responses, not entire inboxes or photo libraries, but the move raises questions about data privacy and the extent of AI integration into personal data. <a href="https://app.daily.dev/posts/Kw6SK0Mqr" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>Rails 8 remains the MVP champion</strong>: Ruby on Rails 8 continues to be a top choice for Minimum Viable Product (MVP) development in 2026, offering built-in authentication, background jobs (Solid Queue), caching (Solid Cache), and WebSocket support (Solid Cable) without external dependencies, making it ideal for rapid iteration and scalability. <a href="https://app.daily.dev/posts/w2WfkMTD2" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>JetBrains IDEs get OpenAI Codex integration</strong>: JetBrains IDEs, including IntelliJ, PyCharm, WebStorm, and Rider, now natively integrate OpenAI Codex (powered by GPT-5.2-Codex) starting with version 2025.3, offering configurable autonomy levels and ‘reasoning budgets’ for AI-assisted coding tasks like refactoring and security vulnerability detection. <a href="https://app.daily.dev/posts/uAvdGTErT" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>GitHub releases multi-platform SDK for Copilot agents</strong>: GitHub has launched a multi-platform SDK for Python, TypeScript, Go, and .NET, enabling developers to integrate Copilot’s agentic workflows, including orchestration, planning, and programmatic file edits, into their applications, though it’s currently in technical preview and requires a Copilot subscription. <a href="https://app.daily.dev/feed-by-ids?id=GXXnGjqMp&amp;id=GuG7K2ryZ" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Feast joins PyTorch ecosystem for feature management</strong>: Feast, an open-source feature store, has integrated with the PyTorch Ecosystem to provide a unified API for consistent feature management across AI model training and deployment, aiming to eliminate training-serving skew and accelerate deployment pipelines. <a href="https://app.daily.dev/posts/dPG1hvjH4" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>New Pascal IDE supports non-visual components</strong>: Quartex Pascal IDE now supports non-visual components and datamodules, allowing developers to drag-and-drop elements for managing database connections and communication protocols, particularly for NodeJS and Deno server environments, significantly reducing server infrastructure scaffolding time. <a href="https://app.daily.dev/posts/sBj3gNusJ" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Rust 1.93.0 improves DNS resolver for static Linux binaries</strong>: Rust 1.93.0 updates its bundled musl libc to 1.2.5, enhancing the DNS resolver for static Linux binaries, improving reliability with large DNS records and recursive name servers, and stabilizing 23 APIs including MaybeUninit slice operations. <a href="https://app.daily.dev/posts/wF6DW7Hp4" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Vercel Sandboxes get filesystem snapshots</strong>: Vercel Sandboxes now supports filesystem snapshots, allowing users to capture and reuse the complete state of a running sandbox environment, streamlining workflows for parallel testing and experimentation by eliminating repetitive setup tasks. <a href="https://app.daily.dev/posts/Uv3Zx9otG" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>AI agents will revolutionize IT operations, but don’t let them touch production</strong>: Apparently, AI agents are the future of IT operations, capable of generating playbooks and Dockerfiles faster than you can say ‘technical debt.’ But wait, the experts (unnamed, of course) warn against letting these ‘non-deterministic’ marvels touch live production systems. So, they’re great for exploratory tasks, but when it comes to actual work, we still need ‘human review, testing, and reproducibility.’ Sounds like we’re just shifting the ‘cowboy sysadmin’ problem to a new, more expensive, and equally unpredictable ‘AI cowboy’ problem. <a href="https://app.daily.dev/posts/U8PYVjc73" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI-driven security agents will dismiss 90% of your vulnerabilities</strong>: ZEST Security is here to save your overwhelmed security team with AI Sweeper Agents that promise to identify which vulnerabilities are ‘truly exploitable.’ They claim over 90% of high-critical vulnerabilities are dismissible, and early adopters have already dismissed 11 million in six months. Because, you know, traditional scanners are just too dumb to understand ‘context.’ No independent verification, of course, but who needs that when you have AI telling you everything’s fine? <a href="https://app.daily.dev/posts/LDTl5G8Jm" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Yann LeCun declares LLMs dead, long live world models</strong>: Turing Award winner Yann LeCun, fresh off his Meta departure, has launched AMI Labs to build ‘world models,’ because apparently, large language models are too dumb to understand the ‘physical world,’ reason, or plan. His new venture, based in Paris, will use ‘joint embedding predictive architecture’ (JEPA) trained on video, audio, and sensor data. It’s an ‘open-source alternative’ to those pesky US and Chinese AI giants. So, while everyone else is still trying to figure out LLMs, LeCun is already on to the next big thing that will definitely, maybe, probably, eventually work. <a href="https://app.daily.dev/feed-by-ids?id=bsXOFS9xq&amp;id=wOdmxGV2q" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
