Theo Browne reviews Anthropic's account of making Claude's web and desktop experiences faster with a Slack-based agent workflow. The discussion focuses on measuring real user journeys, distinguishing client and server work, and supplying agents with benchmarks they can improve rather than vague requests to make an application faster.
Theo Browne tests the public interface and reports problems with conversation persistence and stale sidebar data. He treats aggressive caching and poorly chosen optimization targets as possible explanations, not proven diagnoses of Anthropic's private code. Deterministic instruction counts and rendering metrics are useful only when they correlate with user-visible latency and preserve correct behavior.
Theo Browne highlights benchmarks, regression ratchets, feature flags, incremental rollout and human ownership as safeguards. Detailed examples include syntax highlighting, layout shifts, static composers and frame-budget tests. His broader conclusion is that agents need both measurement tools and engineering judgment, with attention to whether tiny performance wins justify added maintenance complexity.
Watch on YouTube
