Autonomous Development: An Expirement

Overview

I built a controlled bake-off that runs the same long-horizon, wide-scope deliverable through different agentic coding setups, holding the task constant so the only variable is the harness and the workflow layered on top.
The experiment crosses two axes: harness (Claude Code vs. Codex) and workflow methodology (/Ultracode vs. Compound Engineering .

The deliverable under test was creating health dashboard & team of AI health Agents based on my data.

Tech stack