<p><strong>TLDR:</strong> Anthropic added a system-prompt blacklist blocking “OpenClaw” and “OpenCode” by name, triggering immediate backlash from developers who rely on those tools with Claude subscriptions. Cursor 3 shipped a ground-up redesign that demotes the traditional IDE to a fallback and puts agent orchestration front and center. Alibaba’s Qwen 3.6 Plus is now free via OpenRouter with a 1M token context window, and Gemma 4 is running fully offline on phones at 40 tokens per second.</p>
<hr>
<h2>Anthropic blacklists OpenClaw and OpenCode by name</h2>
<p>Anthropic quietly added a system-prompt rule that causes Claude to refuse requests when it detects “OpenClaw” or “OpenCode” in the prompt. The move appears tied to GPU capacity pressure and third-party apps bypassing prefix caching economics, but the implementation — a literal string blacklist — landed badly. Several developers immediately switched their OpenClaw setups to GPT or swapped in Kimi K2.5 and MiniMax M2.5, reporting no meaningful quality difference. One developer documented Claude Code becoming outright unusable as a side effect. The community reaction is less about the policy rationale and more about how opaque and blunt the enforcement was.</p>
<h2>Cursor 3 makes the agent console the primary interface</h2>
<p>Cursor 3 (codename Glass) is a full rewrite that treats the IDE as a fallback rather than the default surface. The new interface centers on a multi-repo agent dashboard with a unified sidebar pulling in agents from mobile, web, Slack, GitHub, and Linear. It also ships Composer 2 (their in-house model), Automations for event-triggered agents, Cloud Handoff for session portability, and self-hosted cloud agents. This puts Cursor in direct architectural disagreement with the rest of the field: Anthropic went terminal-first, OpenAI went multi-surface, and Cursor is betting the agent console becomes where developers actually live. The release comes with Claude Code reportedly running at a $2.5B annual run rate, so the competitive pressure is real.</p>
<h2>Qwen 3.6 Plus: 1M context, free via OpenRouter</h2>
<p>Alibaba released Qwen 3.6 Plus, a closed-weights model built for agentic coding and repo-level tasks. The 1M token context window is the practical differentiator — it makes whole-repo tasks tractable where smaller context models choke. Access is free through Qwen Code (1,000 requests/day via OAuth) and OpenRouter. Alibaba also served 1.4 trillion tokens in a single day on its free tier, 7.3x the volume of the second-place model. That’s not organic growth — it’s a land grab.</p>
<h2>Gemma 4 runs offline on phones</h2>
<p>Google’s Gemma 4 is now running fully on-device on iPhones at 40 tokens per second with no cloud dependency, and it’s hit the top spot on Hugging Face. The model family ranges from a 1.5GB variant that runs on a Raspberry Pi up to a 26B sparse-activation model. Google AI Edge Gallery (iOS/Android) adds Thinking Mode and Agent Skills on top. For most daily tasks, this is a credible offline option — which makes the $200/month subscription math harder to justify for a lot of use cases.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Claude Code source leak:</strong> Anthropic accidentally exposed 512,000+ lines of Claude Code source on March 31 with no reverse engineering required, and trojanized copies have since appeared on GitHub.</li>
<li><strong>OpenAI $122B raise:</strong> OpenAI closed at an $852B valuation with $2B monthly revenue, 900M weekly ChatGPT users, and Codex growing 70% month-over-month to 2M weekly users.</li>
<li><strong>Claude Code benchmark across 13 languages:</strong> A 600-run benchmark found dynamic languages (Ruby, Python, JS) are 1.4–2.6x cheaper than statically typed ones when using Claude Opus 4.6; adding type checkers imposed 1.6–3.2x slowdowns from increased thinking-token usage.</li>
<li><strong>SWECI benchmark:</strong> A new long-horizon maintenance benchmark shows even the strongest coding agents break previously working code more than 75% of the time across 100 real-world tasks.</li>
<li><strong>ARC-AGI-3 leaderboard:</strong> Grok 4.2 scored 0% ten days after release; GPT-5.4 is currently leading.</li>
<li><strong>Alibaba PrismAudio:</strong> Tongyi lab released a video-to-audio model under MIT license that generates synchronized sound from video in 0.63 seconds.</li>
<li><strong>AWS DevOps Agent:</strong> AWS launched an autonomous SRE agent integrating with CloudWatch, Datadog, Grafana, and PagerDuty; United Airlines and T-Mobile report 75–77% reductions in mean time to resolution.</li>
<li><strong>Pinterest MCP:</strong> Pinterest shipped a production-ready MCP ecosystem with human-in-the-loop approvals for security and governance.</li>
<li><strong>Karpathy AutoResearch:</strong> The autoresearch pattern (one modifiable file, locked eval, git-based commit/revert loop) is getting traction beyond ML — practical demos show portfolio load time cut from 50ms to 25ms in under 4 minutes.</li>
<li><strong>North Korean VSCode/Cursor exploit:</strong> Hackers reportedly exploited a VSCode/Cursor vulnerability to steal $285M, posing as trading entities.</li>
<li><strong>eGPU driver for Mac:</strong> Apple approved AMD and NVIDIA eGPU drivers for Thunderbolt/USB4 on Mac, enabling local LLM inference on external GPUs without SIP bypass.</li>
<li><strong>Gemma 4 on LM Studio:</strong> Running Gemma 4 26B locally on an M4 Pro MacBook Pro with 48GB RAM hits 51 tokens/sec at 48K context with a 17.99GB footprint.</li>
</ul>


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

Agentic Digest

A roundup of major AI coding tool developments: Anthropic added a system-prompt blacklist blocking OpenClaw and OpenCode by name, drawing developer backlash. Cursor 3 (codename Glass) is a full rewrite that demotes the IDE in favor of an agent console as the primary interface, with multi-repo dashboards and event-triggered automations. Alibaba's Qwen 3.6 Plus offers a 1M token context window free via OpenRouter, while Google's Gemma 4 runs fully offline on iPhones at 40 tokens/second. Additional items cover a Claude Code source leak, OpenAI's $122B raise at an $852B valuation, a benchmark showing dynamic languages are cheaper with Claude Opus, the SWECI maintenance benchmark revealing agents break working code 75%+ of the time, and an AWS autonomous SRE agent reporting 75-77% MTTR reductions.

Anthropic blacklists OpenClaw in Claude's system prompt, Cursor 3 bets on agent consoles over IDEs