Coding agents raise output, but delivery and evaluation remain bottlenecks
Early firm-level evidence reports more code and pull requests without a significant rise in resolved work, and 49% more review time. New evaluation work includes Accelerate Bench’s 32 end-to-end CPU optimization tasks and proof-oriented benchmark design, alongside the 1,661-case temporal repository benchmark. Evidence that agents improve real delivery remains limited.
Sources (3)
Updated Nov 17, 2026