Chollet Spotlights Test-Time Scaling's Hidden Breadth

The key shift is from making one trajectory smarter to searching more of the solution space. Breadth can unlock better answers, but only when agents are diverse and a strong verifier can separate signal from noise.

• –Larger candidate pools can improve both solution quality and error detection through cross-response comparison. [PMLR research](https://proceedings.mlr.press/v267/zhao25a.html)
• –Agent systems face a real tradeoff: sequential scaling can under-explore, while parallel scaling increases repeated tool calls and cost. [Google Research](https://research.google/pubs/cost-effective-agent-test-time-scaling/)
• –Multi-agent coordination helps parallelizable tasks but can hurt sequential ones, so task structure should determine the depth-to-breadth allocation. [Google Research](https://research.google/blog/towards-a-science-of-scaling-agent-systems-when-and-why-agent-systems-work/)
• –Developers will need better diversity controls, verifiers, aggregation, and budget-aware stopping rules—not simply more agents. [Agent scaling research](https://arxiv.org/abs/2506.12928)