<h2>Must know</h2>
<ul>
<li><strong>AI agents are here, but still need human oversight</strong>: Anthropic’s engineering team is struggling to design technical evaluations that their advanced Claude models (Opus 4 and 4.5) can’t ace, even resorting to Zachtronics-inspired puzzles to find tasks where human reasoning still holds an edge. This comes as Mercor’s Apex-Agents benchmark reveals leading AI models like Gemini 3 Flash and GPT-5.2 achieve only 24% accuracy on complex white-collar tasks, highlighting a significant gap in workplace readiness. Meanwhile, eBay is banning unauthorized AI shopping agents, signaling growing concerns about autonomous AI in e-commerce, as reported by Anthropic and Mercor. <a href="https://app.daily.dev/feed-by-ids?id=vA7gH0M1s&amp;id=OsmrZ9UpM&amp;id=eNz11CwvB" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>OpenAI Codex integrated into JetBrains IDEs</strong>: OpenAI Codex, powered by GPT-5.2-Codex, is now natively integrated into JetBrains IDEs (IntelliJ, PyCharm, WebStorm, Rider) starting with version 2025.3. This allows developers to access the AI agent for tasks like refactoring, code migrations, and security vulnerability detection, with configurable autonomy levels and ‘reasoning budgets.’ Authentication options include JetBrains AI subscription, ChatGPT account, or an OpenAI API key, as announced by JetBrains. <a href="https://app.daily.dev/posts/uAvdGTErT" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Microsoft launches unified speech-to-text model</strong>: Microsoft has released VibeVoice-ASR, a unified speech-to-text model capable of processing up to 60-minute audio files in a single pass. It offers automatic speech recognition, speaker diarization, and timestamping, producing structured transcripts that detail who spoke, when, and what was said. The model supports custom hotwords and targets meeting and conversational scenarios, according to Microsoft. <a href="https://app.daily.dev/posts/zq6H3l6U4" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>Feast joins PyTorch ecosystem for feature management</strong>: Feast, an open-source feature store, has joined the PyTorch Ecosystem to provide a unified API for consistent feature management across AI model training and deployment, addressing training-serving skew. <a href="https://app.daily.dev/posts/dPG1hvjH4" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>TypeScript introduces runtime type stripping for no-build workflows</strong>: TypeScript 5.8 and Node.js 22.6 now support type stripping, allowing type annotations to be removed at runtime for server-side ‘no-build’ workflows, though it requires avoiding TypeScript enums and namespaces. <a href="https://app.daily.dev/posts/gB9bktBA8" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>New AI-driven software development methodology</strong>: AI-DLC is a new software development methodology where AI acts as the primary orchestrator, initiating workflows, decomposing tasks, and generating code, with human validation at key decision points. <a href="https://app.daily.dev/posts/SNNDpvqm6" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>GitHub releases multi-platform SDK for Copilot agent integration</strong>: GitHub has launched a multi-platform SDK for Python, TypeScript, Go, and .NET, enabling integration of Copilot’s agentic workflows, including orchestration, planning, and programmatic file edits, into various applications. <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>BoxLang AI v2 released for JVM AI development</strong>: BoxLang AI v2, a major update for the JVM AI framework, introduces autonomous AI agents with memory, tool usage, integration with 12+ AI providers, and multi-tenant vector memory management with RAG. <a href="https://app.daily.dev/posts/rm0dNUEhj" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Datadog launches automatic retention filters for RUM and APM data</strong>: Datadog released cross-product retention filters that automatically retain Application Performance Monitoring (APM) traces linked to Real User Monitoring (RUM) sessions, unifying frontend and backend observability data for troubleshooting. <a href="https://app.daily.dev/posts/fs5040Sq6" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Vercel Sandboxes introduces filesystem snapshot feature</strong>: Vercel Sandboxes now supports filesystem snapshots, allowing users to capture and launch new sandbox environments from saved states, streamlining CI/CD workflows, testing, and rapid prototyping. <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-generated PowerShell backdoors: because human hackers weren’t efficient enough</strong>: Check Point Research found that the North Korean threat group KONNI is now using AI to generate PowerShell backdoors, complete with obfuscation and instructional comments, targeting software developers and blockchain teams. Apparently, even state-sponsored hackers are outsourcing their malware development to AI. Because why bother with human ingenuity when you can just prompt an LLM to write your next zero-day? This is definitely going to make our jobs easier, right? <a href="https://app.daily.dev/posts/wO1HB29Wl" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI calendar startup Blockit: because Calendly was just too much manual effort</strong>: Blockit, an AI-powered scheduling platform, just raised $5 million to automate calendar scheduling by having AI agents ‘negotiate’ between users’ calendars and learn preferences. It’s already used by 200 companies, and they’re charging $1,000-$5,000 a year. Because clearly, the biggest problem in enterprise productivity was the sheer mental strain of clicking a link to book a meeting. I’m sure these AI agents will never accidentally schedule you for three meetings at once or accept a meeting with ‘urgent’ in the subject line when you’re on vacation. <a href="https://app.daily.dev/posts/R9REfjqEf" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI-driven QA analytics: because 13 years of data guarantees insight, not just noise</strong>: Testlio launched LeoInsights, an AI-based QA analysis solution trained on 13 years and 2.6 million test cases of their own data. It promises executive summaries, a value calculator, and sentiment analysis to convert QA metrics into ‘actionable business insights.’ Because nothing says ‘actionable’ like an AI summarizing 13 years of bug reports without any independent validation. I’m sure this will finally solve all our quality problems, or at least give us a fancy dashboard to look at while the bugs pile up. <a href="https://app.daily.dev/posts/xZI6Dgtwy" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
