DEV TOOLS Signal 185
The Orchestrator's Tax
Engineers building multi-agent AI workflows need to treat the orchestrator's context window as a scarce resource where every token competes for attention. Status-checking mechanisms that dump full subagent transcripts into the main thread can impose costs that dwarf the duplication overhead of running multiple agents, making deliberate delegation rules essential rather than optional.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
In a real Claude Code session, the orchestrator's status-check calls pulled tens of thousands of tokens of raw subagent transcripts into the main context, introducing a large avoidable cost that appeared larger than the duplication overhead of running four agents.
The real value of subagents is protecting the orchestrator's working memory by offloading reasoning it doesn't need to hold, not the wall-clock time saved through parallel execution.
Effective multi-agent work requires giving the orchestrator explicit ground rules for when and how to delegate, turning ad-hoc session behavior into standing rules rather than relying on the orchestrator to self-manage its context.
THE CLUSTER