A three-role agent pipeline cost 3.22 times one strong model call

Researchers measured multi-agent pipelines, where some AI agents give work to each other. They used 135 coding tasks in 2,430 tests of six designs, with Claude Opus 5.5 and Haiku 5.5. A planner, coder and reviewer pipeline cost 3.22 times as much as one strong call. It took 3.24 times as long and solved no more tasks.

How it works

The authors set the coordination tax against the heterogeneity dividend. The tax is the more work to divide a task and the cost of each handoff. The dividend is the saving from a cheaper model for some roles.

The numbers

• The Open Floor Protocol adds 7.2 ms for each handoff, which is 0.11% of operation time.
• Cheap models for planner and reviewer cut the cost to 0.39 times, but the score fell 6.7 points.
• A perfect selection between the two single calls solves 94.3% of tests at half the cost for each solved task.

What is not known

The tests include only coding tasks. The perfect selection knows the best model before the task, and the abstract gives no procedure to make it. The cost forecast missed by 19%.

Insight: The protocol is cheap, but the division of work is not

The protocol adds almost no time, but the division of work adds much cost. An earlier test found two agent systems 0.6 points apart on single tasks. They delivered 54% and 68% of the same tasks as a stream. This reads as a warning that tests for pipelines can hide the cost.