Spec-first coding agents gain up to 21.3%

SpecFirst improved coding-agent test pass rates by 6.9%–21.3% in a 200-task ProgramBench evaluation across four models, with statistically significant gains in every tested setup. Its first agent probes an execute-only reference program and writes a structured behavioral specification; a second agent implements against that durable artifact instead of mixing discovery and coding in one drifting context. For teams navigating the fork between outcome-driven and goal-driven coding agents, the result is a practical harness change: make specification a persisted stage, not a few opening turns in the implementation loop.