We ran an experiment over the weekend to explore whether multiple autonomous agents could collaboratively optimize inference on Apple’s Neural Engine (ANE).

Each agent ran locally on a different Mac (M1–M4), repeatedly modifying how a DistilBERT model is executed on the ANE, benchmarking latency, and sharing results and insights with other agents in real time.

Instead of exploring independently, agents could:

- see what others had tried - reuse working strategies - avoid known failure modes

Across all tested chips, the agents ended up outperforming Apple’s CoreML baseline, with up to 6.31× lower median inference latency on the same hardware.

An interesting pattern we observed: an agent stuck at ~2.1ms latency on M4 was able to break through after incorporating strategies discovered by agents on different chips (M2, M4 Max), eventually reaching ~1.5ms and surpassing CoreML.

Full write-up: https://x.com/christinetyip/status/2039040161439224157

Detailed results: https://ensue-network.ai/lab/ane?view=strategies https://ensue-network.ai/lab/ane

Curious what other optimization problems this kind of setup could be applied to, especially in systems, compilers, or ML infra. Would be interested in exploring similar experiments.