@inproceedings{raptors2026-c057,
  title     = {{The Coordination Layer: How Multi-Agent AI Actually Works in Banks}},
  author    = {Martinovych, Pavlo},
  booktitle = {Proceedings of the Raptors Conference},
  volume    = {1},
  year      = {2026},
  publisher = {Hackathon Raptors},
  address   = {London, United Kingdom},
  doi       = {10.68495/rc.2026.C057},
  url       = {https://doi.org/10.68495/rc.2026.C057},
  abstract  = {Most multi-agent AI deployments in financial services are built as collections of independent agents, one per use case, each carrying its own logic, its own policy assumptions and its own integration with core systems. At scale inside regulated institutions that architecture fails in predictable ways: policy drift between agents, audit gaps at handoff points, and conflicting actions taken against the same core system. Drawing on production deployments across more than 140 financial institutions, this paper argues that the missing component is a shared coordination layer, and shows what that layer looks like in a system that has to survive compliance review rather than only work. The practical question it answers is one of placement: what belongs in the coordination layer and what stays at the agent boundary, a division that determines whether adding the next agent makes the estate more capable or less governable. Readers take away architectural patterns for coordinating multiple agents in production.}
}