← All articles
Unit economics • Field note

Split one agent into five and the token bill can triple

Breaking one agent into a team of specialists is good design. It also adds costs that never show up as an error. You only see them if you meter each step on its own.

David Soden • 5 min read • 9 November 2026
A row of utility meters mounted on a blue industrial wall, one meter per line
One meter per line is how a building finds the leak. An agent pipeline needs the same.

What happened

Towards Data Science ran a post-mortem titled The 3x Token Bill We Didn't See Coming. A team took a working single-agent system and rebuilt it as several specialist agents under a supervisor. The design was sound. Their token usage roughly tripled.

They had assumed cost would grow with the amount of work. It grew with the amount of coordination. Every sub-agent carried its own instructions and its own copy of the context, and each handoff passed along more than the next step needed.

The biggest leak was retries nobody saw. When a later step failed a validation check, the pipeline ran the earlier agents again without telling anyone, and every retry used the most expensive model whatever had gone wrong. No error appeared anywhere. Capping the number of retries barely helped.

Three changes fixed most of it. They sent mechanical steps to smaller, cheaper models and kept the large model for real reasoning. They trimmed what passed between agents down to the fields each one used. And they ran independent branches in parallel, so a failure in one branch stopped forcing a rerun of the others. The team reports about 40% fewer tokens and latency down by roughly half, while admitting they can't say exactly which change did how much.

Why this matters to your business

Most companies watch AI spend as one number on one invoice. That works while the system is a single model answering a single question. Once the work is split across steps, the invoice tells you the total and nothing about where it went.

This failure stays quiet in a way that should bother a finance lead. A retry loop that triples the cost of a ticket looks, from the outside, like a ticket that got resolved. The dashboards stay green while the margin on every automated conversation is a third of what the business case assumed. You find out at budget review, months after the change that caused it.

The team's own measurement gap matters too. They saved real money and still could not say which fix earned it. If the people who built the pipeline can't attribute cost by step, nobody further up the org chart can either.

Interior of an automated bottling plant with conveyors carrying bottles between machines
A production line is costed station by station. An agent pipeline rarely is.

Why this is a CX-Builder use case

Splitting work into specialists is still the right call for most support operations. One agent triages, another looks up the order, a third drafts the reply. That structure is easier to test and easier to fix than one giant agent. The lesson is to build it with the meters in from day one.

In CX-Builder each specialist is a separate step in a visual flow, and each step picks its own model. Classifying a ticket or pulling an order number goes to a small, cheap model. Writing a careful answer to an angry customer goes to the large one. Runs can be traced step by step, so you can see how many tokens each agent used and how often it retried. Because the platform is self-hosted, the flows and their data stay on your servers.

A set of open-ended wrenches laid out in a row from largest to smallest on a brown board
Nobody reaches for the biggest wrench for every bolt. Models work the same way.

What this looks like if you build it

A builder lays the pipeline out as an agentflow with a supervisor and three or four worker agents, each set to the cheapest model that passes its test cases. Handoffs use structured output, so the next agent receives named fields instead of the whole conversation. A condition node checks each result before it moves on, and iteration limits stop a failure from looping back through the expensive part of the flow; past the limit, the case goes to a person. Tracing goes on before the first live ticket, and the per-step token counts get a weekly look.

The takeaway

Before you approve a multi-agent rollout, ask for one number per step: tokens per finished ticket, broken out by agent, with retries counted separately. If the team can't produce it, the pipeline isn't ready for volume. Then check which steps run on your most expensive model and ask whether each one needs it.

All articles Install CX-Builder View on GitHub