Technical Insights

I ran a manager-worker agent team on 67 real tasks. A single session would have done the work better. The failure modes are structural.

Agent memory should be files in git, next to the work they describe, not rows in a vendor database.

A gate an agent waves itself through is not a gate. Quality gates for agent work have to be decisions against artifacts.

The unit of work is no longer a prompt. It is a loop with a driver, a body, exit conditions, and failure modes.

Production got cheap. The bottleneck moved: holding fragments together until the work is actually done.

After Véra on a 1st-gen Scribe (5.19.5): linux/arm on /mnt/us, Wi-Fi SSH, git, a GL.iNet lab tap, and Notes Sync to push stock notebooks to a git remote I control.

Nine months of running Festival on real work, measured from the logs.

How I used Festival to build the Obey Agent Economy at ETH Denver 2026: eight repos, six chains, a vault that can refuse a trade, and 2nd place at Chainlink Convergence.

A battle-tested workflow for using Claude Code to manage large-scale projects with parallel AI agents working like a distributed engineering team

I delivered $1 million worth of software per week for seven straight weeks using AI. Traditional COCOMO values the 219,400 lines of executable code at \(7.8M. My actual cost: \)636 in AI subscriptions. This is how I did it.

How developers are using git worktrees with Claude Code to run multiple AI agents in parallel—early patterns, real implementations, and practical setup guides.

Tech layoffs between 2022 and 2025 weren’t caused by AI automation—they were triggered by a forgotten clause in the U.S. tax code: the expiration of full R&D expensing under IRC §174. Learn what really happened, why it matters, and how to respond.