Gated Human-in-the-Loop Multi-Agent Architecture Cuts Failure Severity 1.58 to 1.16 on Economic Theory Tasks
Zhu, Wang and Zhang address multi-agent reliability when no cheap machine-readable correctness signal exists, building pAI-Econ-claude around a shared workspace of inspectable intermediate records, specialized gates that diagnose targeted failure modes and recommend loopbacks without certifying correctness, and human checkpoints holding authority over costly-to-reverse decisions. Across five matched economic-theory tasks against an ungated baseline, two blinded evaluators agreed on all five pairwise rankings, preferring the gated architecture in four; mean failure severity fell from 1.58 to 1.16 and usefulness rose from 2.60 to 3.10. The single negative case matters too — scaffolding compressed an economically important mechanism too aggressively — and the workflow is public at github.com/maxwell2732/pAI-Econ-claude.
↳ Follow the thread