Multi-agent coding systems often fail not due to poor communication, but because commitments made in conversation lack a persistent record. Even when agents communicate effectively, they can duplicate work, misorder tasks, or miss dependencies because no external state tracks who promised what. The author built a Commitment Ledger using only Python's standard library to record agent commitments—like 'I'll build the items file'—as explicit state.

This ledger checks for conflicts, missed work, and rework without requiring databases or APIs. In testing, it prevented duplicate work and dependency failures, but did not fix cases where agents failed to complete tasks or reported unverified work. The ledger improves coordination state but cannot enforce agent follow-through.

The insight stems from the Cooper Bench benchmark by Stanford and SAP Labs, which showed cooperating agent pairs sometimes scored worse than a single agent due to vague messages, incorrect assumptions, and drifted commitments. While coordination is possible—as seen in minority runs where agents negotiated roles or split resources—it is unreliable without a mechanism to persist commitments. The ledger isolates this specific failure mode: commitments existing only in chat, with no way for other agents to verify them later.

The solution is lightweight, deterministic, and focused solely on making commitments visible and checkable.