Multi-agent amplification

Each agent in a pipeline re-sends its own context, adds its own instructions, and may retry independently. Costs therefore grow with the shape of the delegation tree, not the size of the task — and a topology change (one more delegate, one more review pass) can multiply spend without any code looking more expensive in review.

A coordinator delegates to four workers, each of which makes six calls and one review pass. The task the user sees is "one question answered"; the system executed roughly thirty calls. In the July 2026 scan of public agent repositories, the median repository defined 2 agents and the 90th percentile 8 — trees, not single calls, are the normal case.

By attributing every call to both its agent and its run, then comparing run totals against the naive single-agent estimate. The amplification factor is the ratio.

What is multi-agent amplification?

Multi-agent amplification is the compounding of LLM cost that occurs when agents invoke other agents, so one task multiplies into a tree of calls whose total far exceeds what any single node suggests. Each agent re-sends its own context, adds its own instructions, and may retry independently. Cost therefore grows with the shape of the delegation tree rather than the size of the task.

Why are multi-agent systems so expensive?

Because the work is duplicated at every level of delegation: each agent pays for its own context, its own instructions, and its own retries, and a coordinator pays again to read what its delegates returned. Adding one review pass or one more delegate multiplies spend without any individual call looking expensive. In Capsera's July 2026 scan of 133 public agent repositories, the median repository defined 2 agents and the 90th percentile 8 — trees, not single calls, are the normal case.

• Cost per run — Cost per run is the total LLM spend attributed to one execution of an agent task, summed across every call, model, and agent that execution triggered.
• Runaway agent loop — A runaway agent loop is an agent retrying or re-invoking itself without a terminating condition, generating LLM spend continuously until something outside the loop stops it.
• Context regrowth — Context regrowth is the compounding cost of re-sending an entire conversation history with every turn, so total input tokens grow roughly with the square of the conversation length.