Series: Solo + AI Delivery Efficiency — Part 2 of 6.

Two levers, two jumps

A better AI model made me 2.5× faster. Workflow made me 8×. The order matters more than people think.

Lever 1 — the model (May). A single AI assistant lifted my throughput to ~2.5× the human baseline. Real, but capped. One agent, one thing at a time.

Lever 2 — the workflow (June). I stacked an AI work loop on top of git worktrees, so multiple agents run in parallel, each on its own branch, with every change gated through a pull request before merge. That's what took it from 2.5× to ~8×.

This matches what the large-scale research keeps finding: the model alone moves the needle surprisingly little. An analysis of 121,000 developers found AI productivity gains have largely stalled around 10% — because most teams bolt AI onto an unchanged process. The leverage shows up when you restructure the workflow around it.

The flat plan matters here for a non-obvious reason: it removes the rate-limit ceiling that would otherwise cap how many agents I can run at once. You can't parallelise agents if you're rationing tokens.

The model gives you a faster worker. The workflow gives you a team of them.

Sources

Figures are paraphrased; verify against the originals before quoting.

• DX (Laura Tacho), Measuring Developer Productivity & AI Impact — 121,000+ developers, ~10% gains: shiftmag.dev

Previous: Part 1 — 133 Days, One Operator · Next: Part 3 — The Cost Inversion