New from Google, DeepMind and Berkeley,

This is a good example of why better agent architecture matters:

ArchAgent v2 shows a useful pattern for AI discovery: when a problem is too large to search at once, split it up and make the agent obey the same constraints as the final system.

Here, the problem is CPU cache prefetching: predicting which data the processor will need next.

Searching L1D, L2, and LLC together was too slow and messy, so ArchAgent v2 optimized the cache levels in stages, kept the best designs, then refined them together.

It also rejected candidates that exceeded the actual storage budgets.

With that setup, it beat the human-designed DPC4 champion.

Its final design delivered a 3.8% IPC improvement over the competition baseline and edged BertiGO by 0.3%. On low-bandwidth single-core systems, it reached 4.6% versus BertiGO's 2.6%.

Multi-core performance still lagged because those simulations make search much slower.

The takeaway for AI builders: better discovery agents need better search structure and real engineering constraints, not just a stronger model.

– arxiv. org/abs/2608.09874

Title: "ArchAgent v2: A Case Study with the Data Prefetching Championship"