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: the pm-skills repo publishes skill files, each one a shareable unit, that plug into an agent's slots so capabilities are assembled and swapped rather than rebuilt.
A wave of 'agent skills' repos is exploding on GitHub: phuryn's PM Skills Marketplace alone bundles 100+ ready-made skills, and Anthropic just shipped its own official skills repo. The pattern matters more than any single repo — reusable, shareable skills are becoming the app store for AI agents, so you assemble capabilities instead of re-prompting every workflow from scratch. Whoever's skills become the default will shape how teams build, the way app stores shaped mobile. Audit the prompts and workflows your team rebuilds over and over, package them as skills now, and standardize on a format before someone else's becomes the default.
HOW TO READ THIS Read top to bottom: DeepMind ships DiffusionGemma, which denoises all tokens in parallel instead of the transformer's one-token-at-a-time decoding.
The most serious non-transformer LLM yet generates text by diffusion instead of one token at a time — if it holds up, the first real crack in transformer dominance and a faster, more parallel path forward.
HOW TO READ THIS Read top to bottom: the repo's code, Claude Code scanning it, 20 checked vulnerability classes, then the shipped result.
An open-source tool turns Claude Code into an autonomous bug-bounty hunter — recon, 20 vuln classes, and reports from the terminal — a vivid sign agents are moving from writing code to probing it for holes.
HOW TO READ THIS Read top to bottom: a survey of consumers, products stamped with AI tags, the tag triggering rejection, then the 60% who tuned out.
A WordPress VIP survey finds most US consumers are put off by 'AI' in brand messaging — the label you're betting on may be quietly costing you trust.
Open-source RAG engine now fusing retrieval with agent capabilities — a production-grade alternative to rolling your own.
Open-source LLM observability, evals, and prompt management — the instrumentation layer most agent stacks are still missing.
ByteDance's open multimodal agent stack for driving real desktop UIs — GUI agents you can actually run and inspect.
Official Chrome DevTools exposed as an MCP server, letting coding agents inspect and debug live browser sessions.
A meta-harness for orchestrating multi-agent swarms on Claude — coordination scaffolding for autonomous workflows.
Latent Space GLM-5.2 reportedly takes the top frontend-coding spot — open-weights pressure on the closed labs keeps mounting.
Interconnects Argues AI governance has entered its 'AGI era' — worth a read if you're anywhere near policy or compliance.