multi-agent tax

What it means

the extra tokens, computation, time, complexity, or loss of performance caused by coordinating multiple AI agents instead of having one agent perform the underlying task

Where it comes from

The word tax has long been used figuratively for an unavoidable overhead. Multi-agent tax and coordination tax became more visible in 2026 as developers measured the hidden cost of agent teams, including context duplication, handoffs, review, and integration. Research published that year used collaboration tax for a related, more formal measure of the performance lost when language-model agents must coordinate with one another.

In a sentence

The engineering team measured the multi-agent tax before adding more specialist agents to its coding workflow.

Repeatedly copying the full project context into every subagent made the coordination tax larger than the benefit of parallel work.

Good to know

The multi-agent tax is not evidence that agent teams are always inefficient. Several agents can help when a task divides cleanly into independent parts or benefits from genuine specialization. The tax becomes important when agents repeatedly resend the same context, duplicate work, wait on dependencies, or require a lead agent to reconcile conflicting results. Measure quality, total token use, runtime, and human review across the whole task rather than judging each agent separately. Collaboration tax can refer more narrowly to performance lost through imperfect communication, while multi-agent tax often includes financial and operational overhead as well.