All notes with the tag "agents".
A postmortem of NightLoop: why per-task scoped context degrades frontier coding agents, and why governance should verify outcomes instead of constraining what the model sees.
Claude Code can now write an orchestration script and fan a single prompt out to hundreds of parallel subagents. What it's for, and the one reason not to reach for it.
Andrej Karpathy's argument that the bottleneck in AI-assisted building isn't the model — it's the spec you fail to write. Distilled from a walkthrough of his AISN 2026 talk.
A working path through Codex's five surfaces — install, approval modes, AGENTS.md, and the handoff protocol — distilled from Nate B. Jones' living guide, with the sharp edges marked.