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: your Skills repo feeds a read-only MCP link into the review engine, which posts an attributed comment, now GA July 29.
GitHub made agent skills and MCP servers generally available in Copilot code review on July 29. Drop a SKILL.md into .github/skills and the reviewer applies your team's internal tools and coding standards instead of generic opinions, while MCP pulls context from issue trackers and docs with every tool call forced read-only. The part that matters if you ever have to defend a review decision: comments now carry attribution showing which skill or MCP server produced them, which turns a review agent from an opaque opinion into something you can audit. Start by encoding the three review rules your team repeats most often in PR comments — that is where the accuracy gain comes from, not from a bigger model.
HOW TO READ THIS Read top to bottom: a plan is approved, CLI v2.15.0 ships, the old manual step is skipped, and the plan runs straight through to done.
Kiro CLI v2.15.0 (July 27) added a guided description step to /spec new and now runs approved plans automatically instead of requiring a manual mode switch, while IDE v1.0.242 (July 28) added a Kiro context menu, an "Ask Kiro to Fix" quick fix on errors, and a guided form for building hooks. It is a small release that deletes the last human handoff between planning and execution — the plan you approve is the plan that runs. That raises the stakes on plan review considerably: approval is now the only gate left, so treat it like a merge, not a formality.
HOW TO READ THIS Read top to bottom: a repo ships packaged skill files, MIT-licensed and freshly updated, which plug into an agent in place of a discarded one-off prompt, driving 197K stars.
Matt Pocock's MIT-licensed skills repo — "Skills for Real Engineers. Straight from my .agents directory" — passed 197,000 stars and was updated again this week, packaging TDD, code review, bug diagnosis, architecture work, and merge-conflict resolution as small composable SKILL.md files. The signal is not the star count, it is what people are choosing to share: the unit of reuse has moved from prompts to engineering practice, portable across agents and vendors. If your team's review and debugging conventions live in a wiki nobody opens, this is the format to move them into.
HOW TO READ THIS Read top to bottom: github ships the harness, its memory chip shrinks from 2.3GB to 27.8MB, then many sessions fit in the same space.
jcode is an MIT-licensed Rust coding-agent harness whose own published benchmarks claim roughly 27.8 MB per session against 2.3 GB for Claude Code and 1.75 GB for GitHub Copilot CLI, with support for Claude, OpenAI, and Gemini backends. Those are the project's numbers, not independent ones, so measure on your own workload before trusting the ratio. Still worth a look if you run many concurrent agent sessions on one box — that is the regime where harness overhead, not token cost, becomes the binding constraint.
Trail of Bits' Claude Code skills for security research, vulnerability detection, and audit workflows.
Official spec and SDK for MCP Apps — UIs embedded in AI chatbots, served by MCP servers.
LangChain's batteries-included agent harness — the opposite design bet from jcode.
Figma's own guide to using its MCP server — design context as a first-class agent input.
Open-source TTS positioning itself as state of the art, MIT-adjacent and self-hostable.
Ben's Bites Ben's Bites marks ChatGPT crossing one billion users.
Latent Space Latent Space argues AI agents are reviving the semantic web — ontologies as the shared context layer.
The Sequence The Sequence's contrarian read: the age of research is overrated, AI engineering is winning.
SemiAnalysis SemiAnalysis on the wild west of modular "LEGO" datacenter buildouts.