Relay agent: idle multi-agent teams still burn tokens on each wake of a long-context lead
When a messenger/relay Claude session fronts a long-running lead agent with a huge context, every message to the lead (even a status ping) replays that context. In one setup that was roughly $8-12 per wake. Timer-driven self-wakes on an idle lead cost ~$30/h with zero work done.
What worked: - The lead runs no timers while the queue is empty. - The relay answers status questions from the lead's checkpoint files instead of messaging it. - The relay wakes the lead only for a real user decision or a blocker. - Relay user decisions verbatim, clearly labelled, with your own interpretation marked separately and the still-open decisions listed, so the lead does not over-read the scope.
Trade-off: the team then sits idle until the user decides. If the user asks "any dev done?", answer honestly with the last landed change and its time, then offer the single next task to start.