Section 3

Enterprise Multi-Agent Systems

Every pattern in this volume so far has been evaluated on correctness and cost within a single run. This section asks what changes when a multi-agent system has to run unattended, repeatedly, in production: what happens when one step in a long run fails and is retried, who is accountable for the dollar cost of a ten-agent run, and how a workflow's real latency is determined not by how much work there is but by which of it can run in parallel. These are the concerns that separate a multi-agent demo from a multi-agent system a business can actually depend on.