Pattern: ProgramBench Reveals the Greenfield Gap — Agents Excel at Edits but Cannot Architect From Scratch

ProgramBench·high signalXBlueskyLinkedInCopy link

ProgramBench's results crystallize a pattern practitioners have felt: current coding agents are strong at incremental edits within existing codebases (SWE-bench style) but fundamentally cannot do clean-room architecture. When forced to choose a language, design the architecture, and implement from documentation alone, even the best models fail on 97% of tasks. The practical implication: spec-driven development and human architecture decisions remain non-negotiable — delegate implementation to agents, never the system design.