Developer ditches agent-to-agent messaging in favor of shared claims ledger

A developer running seven AI agents over the same domain abandoned the conventional coordinator-and-conversation architecture after it broke down under real workloads. Instead of agents messaging each other, each agent now writes structured claims to a single shared record, with no agent able to read another's reasoning or invoke another directly. Each agent is assigned a strict domain of authority, such as verification, design, or capital, and may only submit claims within that jurisdiction. When agents disagree, a rules-based reconciliation system resolves conflicts using evidence grades rather than agent rank, and losing claims are demoted rather than deleted to keep disagreements auditable. The author argues this flat coordination model eliminates quadratic token costs, prevents error propagation across agents, and makes system decisions explainable to auditors weeks after the fact.