多智能体编排层性能超越Claude Code和Codex

So much recent work and research papers points to the same thing: the "harness" is becoming the real capability layer.

@Offloop 's 4-person team demonstrated a multi-agent harness outperforming Claude Code and Codex on GDPval benchmarks, targeting $2.4T in US knowledge work.

• The team scored 84.9 at $1.65 per task. - Opus 4.8 inside Claude Code scored 82.4, and GPT 5.6 Sol inside Codex scored 83.3, costing far more per task, $14.38 and $5.20

GDPval measures how well AI handles real work across 44 occupations and 9 major industries. Models get shell access and web browsing, then face blind pairwise comparisons against human experts.

And those tasks map onto US jobs paying roughly $2.4T a year.