Coding agents need speed under load, not single-user benchmarks

Original: Benchmarking inference at scale: coding agents

Source: together.ai ↗