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: one agent branches into more agents, each new spawn passes a classifier check, and the branching recurses down a chain of five levels.
Claude Code now lets a sub-agent spawn its own sub-agents up to five levels deep, with the auto-mode classifier vetting each spawn before it launches (v2.1.172 onward). That quietly turns the CLI from a single assistant into an orchestration layer: a planner agent can delegate to specialist agents without you hand-wiring every hop. The catch is the one that comes with any deep delegation — depth multiplies cost and makes failures harder to trace. Pick one multi-step workflow you currently cram into a single mega-prompt, model it as a planner plus a few specialists, and watch the spawn depth and token spend before you trust it unattended.
HOW TO READ THIS Read top to bottom: Cursor ships the upgrade, review time collapses, then cost and bug-catching both improve.
Bugbot on Composer 2.5 drops review time from ~5 minutes to ~90 seconds, costs ~22% less, and catches ~10% more bugs — cheap enough to run on every PR, not just the scary ones.
HOW TO READ THIS Read top to bottom: GitHub ships the tools, the MCP server adds them, a privacy shield gates blame and commit access, then the agent receives it.
Blame and commit history — the context coding agents most often lack — now ship from GitHub's first-party MCP server (get_file_blame, get_commits), so you stop trusting a random community wrapper for it.
HOW TO READ THIS Read top to bottom: GitHub ships OAuth into the gateway, retiring the one shared token in favor of each user logging in and getting a distinct, checked identity.
Cursor and Claude Code can now authenticate to your MCP gateway with real OAuth/DCR instead of a shared static token — the line between a demo and something you put in front of a team.
LLM knowledge-curation agent that researches a topic and writes a full, cited report — research-agent patterns you can actually read.
Garry Tan's exact Claude Code setup — 23 opinionated agents standing in for CEO, designer, eng manager, and release manager.
Review-first terminal diff viewer built for agentic coders — read the diff before you let the agent's changes land.
An ADE for running a fleet of parallel coding agents, each on your own subscription.
754 structured cybersecurity skills for AI agents, mapped to MITRE ATT&CK, NIST CSF 2.0, and ATLAS.
Interconnects Nathan Lambert's case against banning open-source AI — worth reading before the policy fights land on your stack.
The Sequence This week's roundup leads with a reported $60B Cursor deal — the agentic-IDE land grab is getting expensive.