Incentive-Responsive Agents: A Principal-Agent Framework for Long-Lived Personal AI

Summary

This arXiv preprint argues that long-lived personal AI agents need an incentive layer, formalized as runtime contracts over operating points (model tier, compute, memory, tool authority, etc.). The authors derive principles including a non-consumption invariance, persistence-as-discount theorem, incentive-capability exchange-rate frontier, and Goodhart bound. A 240-run real-LLM experiment found prompt-level incentive framing had no effect (p=0.62), while runtime allocation raised hardest-task success from 0.60 to 1.00 (p=0.008) at 4.8x cost; a 20-seed, 3,000-task simulation showed ROI rising from 7.09 to 16.67 and gaming falling from 0.333 to 0.115.

This is a reviewed summary. Read the complete article at the original source.

Why it matters by industry

• Artificial intelligence

  Relevant to AI model and agent development, particularly for long-lived personal agents, offering a formal incentive framework and empirical evidence on runtime allocation versus prompt engineering.