Hierarchy scaffolds break agent idea plateaus

Autonomous coding agents often plateau in open-ended scientific tasks not because they can’t code, but because they can’t generate novel hypotheses—they lack “research taste.” The talk reveals that the real bottleneck is idea generation, not implementation.

The speaker’s team at Radicate overcame this by explicitly decomposing the problem into a documented hierarchy of subcomponents (data, architecture, loss, metrics) before starting the optimization loop. The agent used this scaffold to traverse and propose improvements at every level, producing radical changes—like switching from 2D to 3D convolutions—that a flat prompt never would.

This technique functions like chain-of-thought applied at the domain level, scaling test-time compute across a structured search space. The trade-off is the upfront effort to create the hierarchy, plus the need for human or adversarial review to validate proposals. Integrating multimodal quality-control agents can further improve the loop.

The evidence is anecdotal (one ML model optimization), and multimodal observation remains weak. Future models may internalize this ability, but today engineering leaders can adopt explicit hierarchy design to sustain agent progress on complex, long-horizon tasks.