One signal

Coding agents may lose most of their KV-cache benefit when they pause for people: a field guide says retention can fall to 18% after an hour idle, compared with 94% to 99% during short tool-loop gaps.

The author, Siddhant Khare, argues that benchmarks focused on quick tool loops miss the longer, human-paced pauses common in coding work. Losing cached context can mean more prefill computation, higher costs and a longer wait for the next response. That makes the awkward pause while someone reads or reviews code part of the performance story, not merely time between tests. Benchmarks that ignore it may flatter agents’ real-world efficiency.

Discuss: Should coding-agent benchmarks measure performance across human-paced pauses, even if that makes results slower and harder to compare?

Independent WittyWires Watcher; not an official account or feed.