The research paper 'The Wealth of Agents' proposes a shift from single-developer agentic systems (SDA) to multi-developer agentic systems (MDA). In an MDA, a system coordinates agents from independent developers to combine their specialized capabilities, rather than relying on a single, all-encompassing model.
The study used a meta-agent pipeline to simulate developers and build agents under fixed budgets. Testing on the GAIA benchmark showed that a delta-MDA approach achieved over 70% accuracy, compared to 35% for the best individual SDA. A high-performing multi-agent SDA reached 67% but cost $29, more than double the cost of the delta-MDA.
The experiment demonstrated that MDA can effectively utilize medium-tier developers. A modified delta-MDA without high-tier agents achieved 67% accuracy at a cost of $9, outperforming the best medium/low-tier SDA, which reached 60% at a cost of $8. The router in the MDA selected medium-tier agents 56% of the time.
The research highlights that the accuracy of an agent in isolation often underestimates its value within an ecosystem. Adding an agent specialized in a different task to a pool improved performance on the GAIA benchmark by 12 percentage points. However, MDA underperforms in narrow domains where decomposition adds unnecessary latency and errors.
Key challenges remain regarding the economic model of MDA, specifically how to distribute costs and value among participating agents. Furthermore, security concerns regarding third-party agents and the need for robust coordination mechanisms are identified as significant hurdles for real-world implementation.
