Blog

From process to microVM: step through one container running many claude subprocesses, toggle isolation layers (cwd, srt, egress proxy, container, gVisor/Firecracker) against seven problems, and see which layers each session pattern needs. Built with Claude Code.

Notes on hosting the Claude Agent SDK: session transcripts, CLAUDE.md memory files and working-directory artifacts all live on the container's disk, and the four session patterns (ephemeral, long-running, hybrid, multi-agent) follow from that.

Step through all 11 frames of MCP's OAuth flow on a sequence diagram: who does what (server side vs. OAuth client side), where PKCE's secret pair travels, and how the auth server's ledger ties the code to the token. Built with Claude Code.

Frame-by-frame playback of a real MCP client: where the thread goes when a coroutine freezes at await, why gather makes two requests overlap, and why the design has to be this way. Built with Claude Code.

Notes on Anthropic's "When AI builds itself": recursive self-improvement is just better tools, and the real limit inside a company is capacity, not capability.

The function is durable, the title probably isn't — and the durable half of the value is the least portable half.

Three things I'm building in myself — a mental model of tech, judging my own judgement, and taste.

Three ideas that all point the same way: in the AI era, understanding is the bottleneck — not typing speed.

Two functions that both "wait two seconds." One freezes the thread; the other lets it keep working. Plus an interactive timeline.

Why I start from the workflow, not the model — and the pattern that keeps working.