Cohere senior researcher Nouha Dziri presented tests of AI agent teams using a lead that delegates work and combines reports in a shared workspace. Pairing the more capable Sol lead with cheaper Luna planners delivered the same result as using Sol throughout, at a lower cost.
The tests used ProgramBench, which asks models to reconstruct a program from its binary and specification so it compiles and passes tests. Restricting workers to exploration achieved the full team's performance while halving costs, suggesting that discovering what to build was the main bottleneck in this setup.
For companies building AI agents, the findings support testing where extra agents help before paying for them throughout a workflow. They do not establish a general cost advantage: another benchmark, Commit Zero, rewarded implementation help when documentation and tests were supplied.