Does AI coding actually work? What the evidence shows
A controlled trial found a 19% slowdown, surveys describe AI as an amplifier, and operators report thousands of agent pull requests a week. Each result measures something different.
Ideas
The loops, work graphs, operating models, and unresolved arguments changing how software gets made.
A controlled trial found a 19% slowdown, surveys describe AI as an amplifier, and operators report thousands of agent pull requests a week. Each result measures something different.
HarnessTax compared 21 model–harness pairs on two public coding-agent benches. Harness choice moved token cost far more than success. Do not inherit a provider default without a comparison.
How to combine three ways of organizing agent work: a production line for one repeatable job, Geoffrey Huntley's Ralph loop, and Steve Yegge's multi-agent work graph.
A personal manifesto about building more software by improving the system around coding agents.