New AI model cuts agent costs by 99% with smart routing
AI agents cut costs 99% by activating only 2 of 8 needed, new paper shows
Researcher Mojtaba Eslami has proposed a novel framework for optimizing multi-agent AI systems by modeling agent selection and communication as a cooperative game. The work, published on arXiv (arXiv:2608.07532), introduces a marginal-value activation rule and greedy router that significantly reduces inefficiencies in current agentic AI architectures.
In synthetic experiments, the proposed system activated only 1.96 agents on average out of 8, achieving 99.5% of brute-force-optimal utility—far surpassing the 38.8% achieved by full broadcast methods. The framework leverages Shapley value estimates to predict which agents are worth contacting before and during execution, while proving approximation guarantees via submodular maximization. Performance remained robust to activation cost variations but dropped to 66% under strong submodularity violations or noisy value estimates.
- Greedy router activates only 1.96 of 8 agents (vs. 38.8% for full broadcast) while achieving 99.5% optimal utility
- Uses Shapley value estimates and submodular maximization with 1/2-approximation guarantees
- Performance drops to 66% under submodularity violations or noisy value estimates
Why It Matters
This work could slash AI agent costs by 99% in real-world deployments, making multi-agent systems far more efficient and scalable.