Lucky timing: I spent the week rebuilding the Fidero site. Every page was spec'd with Fable 5 the day before the US government ordered it suspended, and Anthropic pulled access for everyone. Glad I got the heavy thinking done first. The whole build was agent-driven, and it's becoming a blog series that walks through the process end to end. Anthropic's statement

Context is data: Someone asked how I keep a 400k-token task coherent. My answer's always the same: treat context like a data engineering problem. Context is data. It rots the same way messy data does, and the fixes are the same ones data teams have leaned on forever: be deliberate about what goes in, and feed the agent only what it needs, when it needs it. It's the lens I reach for because my background's all data, but plenty of people get there other ways.

Running out of context? Scope's too big: A common worry on a long build is hitting the context limit. If it's happening on one task, that task is doing too much – break it into smaller units. I had four agents running in parallel in a single day, each in its own session and git worktree (an isolated working copy), and none of them came close. Keeping track of what's been built across sessions is a separate problem, and that's what a memory bank solves. Blog post

On my radar

Local models are good now: I keep seeing the case that local and self-hosted models have crossed into genuinely useful. The fair version: good enough for overnight batch jobs and cost control, still too slow for anything interactive. Blog post

Cursor launched Origin: A GitHub alternative built for a world where agents commit code faster than the old tools assume (they claim 22 commits a second). When machines write most of the commits, those assumptions crack. Cursor

Skills are meant for sharing: The gap in agentic tooling is the non-technical person who can't wrangle CLIs and MCP servers (the plumbing that connects tools to your AI). Portable skills that work across tools and people are the real unlock. Blog post