Dinner on the Economics of Multi-Agent Swarms: Cross-Company Models May Never Learn to Coordinate

HaydnBelfield · x · 2026-11-13

Economist Andrey Fradkin hosted a dinner on the economics of multi-agent swarms, with several notable observations:

There's no good taxonomy or theory yet for when to use swarms vs. single-threaded or scaffold-style agentic systems. Some argued that as models get smarter, swarms' value may reduce to speed and parallelization.
Where swarms do remain useful, it's unclear whether current training techniques let agents learn optimal self-organization, or whether a more Darwinian path — agentic organizations evolving under competition — is likelier.
Theoretically, heterogeneous multi-model swarms benefit from diversity, but models trained separately by different companies may never be trained for optimal mutual coordination — a structural gap in cross-model coordination training.

A rare economics-lens discussion on agent orchestration.

More from AGI Musings