AI that matters, from the architect's desk. Curated and engineered by Saaket Varma, PhD — no hype, just signal.
HOW TO READ THIS Read top to bottom: Anthropic ships Sonnet 5, it becomes the routing hub for products as the new production default while the old default is retired, then joins the Claude 5 family.
Anthropic announced Claude Sonnet 5, the newest model in the Claude 5 family, and developer write-ups are already dissecting what's new. This matters more than a frontier-tier launch would: Sonnet is the tier most production workloads actually run on, so this release resets the default price-performance choice for anyone building with Claude. If you standardized on an older Sonnet because it hit the cost-quality sweet spot, that math just changed. Re-run your eval suite against Sonnet 5 before your next deployment cycle — don't let an outdated model choice become silent technical debt.
HOW TO READ THIS Read top to bottom: Cloudflare sits in front of sites, AI crawlers arrive, each must pay or get blocked, and after Sept 15 the default flips to blocked.
Cloudflare gave AI companies until September 15 to separate the crawlers they use for search from those used for AI training and agents — or be blocked by default across many publisher sites. Because Cloudflare fronts a huge share of the web, this effectively industrializes the pay-for-content fight between publishers and AI labs: what was a lab-by-lab negotiation becomes infrastructure-level enforcement. If you build agents that fetch web content, your access model is changing — audit what your crawlers identify as now, and start treating content access as a real cost line, not a free input.
HOW TO READ THIS Read top to bottom: Stablyai builds Orca, Orca ships as a single IDE window, that window drives four agents at once, and the result is #1 on GitHub with 10,365 stars.
Orca, an 'agent development environment' for running fleets of parallel coding agents on your own subscriptions, is the top trending repo on GitHub — on desktop and mobile. It captures the real workflow shift underway: from babysitting one chat assistant to orchestrating many agents at once, including from your phone. The takeaway isn't this specific tool — it's that fleet orchestration is becoming its own tooling layer, and the engineers who learn to manage parallel agents will outproduce those still working one prompt at a time.
HOW TO READ THIS Read top to bottom: AWS ships a toolkit of servers, skills, and plugins that plug into MCP, making AWS a first-party MCP provider.
AWS released an official, AWS-supported toolkit of MCP servers, skills, and plugins that help AI agents build on AWS, and it's surging on GitHub trending. First-party MCP support from the largest cloud makes agent-driven infrastructure work a sanctioned, supported path instead of a hack. If your team runs on AWS and has been hand-rolling MCP shims or gluing agents to the CLI, this is the moment to consolidate on the supported path — official backing means it will track new services in ways community wrappers can't.
Open-source AI penetration testing tool that finds and helps fix vulnerabilities in your apps.
AllenAI's toolkit for linearizing PDFs into clean text for LLM datasets and training pipelines.
Free AI gateway — one endpoint across 231+ providers, pluggable into Claude Code, Codex, Cursor, and Cline.
Meta's open-source design system, fully customizable and built to be agent-ready.
Turns any technical book PDF into a Claude Code skill you can reference while you work.
Ben's Bites GPT-5.6 has arrived — with enough caveats that the 'but...' in the headline is doing real work.
SemiAnalysis Field notes from enterprise conversations on token spend — how big buyers are actually budgeting LLM costs.
The Sequence Meta's Autodata: models that learn to make their own lessons — a look at self-generated training curricula.