Writing.
Notes from running an agent fleet as coworkers. Multi-agent systems, fleet operation, infrastructure underneath. New pieces shipping as I finish them.
Every long-running agent lives in a detached tmux session named after its job. Six months of fleet operation distilled to one rule.
Token saturation is the only signal that matters. Wall-clock time, idle time, vibe, all decoys.
A SQLite-backed memory layer agents can write to and retrieve from across sessions, providers, and machines.
The recursive structure that lets a single architectural rule survive across sessions, providers, and tier boundaries.
From single-prompt user to fleet operator: the eight components of a personal AI OS, and what shipping each one taught me.
For solo developers and agent-driven sites, vanilla HTML beats React, Next, and Astro. The framework tax is hidden until you remove it.
619 commits in two months. The two load-bearing choices: Rust, because the compiler shortens the agent's iteration loop, and tmux, because agents are long-running processes that have to survive the laptop.
A 200-line Rust binary that does literal substitution between HTML comments. Smaller than every templating system I've used.
Most CMSs treat HTML as a build output. Freedom CMS edits the HTML file directly via contenteditable, autocommits to git. No database, no schema, no build step.
An MCP server over local SQLite, semantic retrieval, lifecycle tiers, conflict detection, and buckets. Published on npm and working across Claude Code, Cursor, and Windsurf.
619 commits and two months in. An honest retrospective: state-in-the-filesystem and folder-as-config held under load; idle polling and handoff races broke.
The first unsupervised run of my agent orchestrator: 686 cycles, 11 merged PRs, 8 self-healed stalls, and 94% idle time. What worked, what broke, and why the failures were the point.
A CLI orchestrator that turns AI agents into long-running workers against a GitHub backlog: a 3-layer master/workspace/project model with a vault-backed feedback loop.
A practical Git workflow for vibe coding with AI: clean branches, atomic commits via git add -p, intent-driven messages, and PR review as the final gate.
Three things I wish someone had told me when I started Webflow eight years ago: use Relume for layouts, follow Client-First for class naming, and switch to REM units.
I ran a fully autonomous AI agent workspace for twenty sprints with no human gates. It shipped 200 pages. The more valuable output was a four-item failure catalog.
When you run agents across fifty-plus projects, the filesystem becomes your bottleneck. A three-type taxonomy — seat, domain, system — cuts orient-cost from O(n) to O(1).