How Anthropic made claude.ai 3x faster in two weeks with Claude merging 3,000+ changes
ClaudeDevs · x · 2026-09-24
Anthropic engineers made claude.ai and the desktop app about 3x faster in a two-week sprint, coordinated entirely through a single Slack channel with Claude in every thread.
Key results (75th percentile): time to typeable page on fresh load dropped from 3.1s to 0.55s, starting a Claude Code session from 0.8s to 0.3s, and loading a Claude Cowork cloud session from 2.6s to 0.73s — an estimated tens of thousands of user-hours saved daily.
How they did it:
- Claude (via Claude Tag beta, an internal model roughly comparable to Opus 5.5) found bottlenecks, built benchmarks, shipped fixes, and watched every deploy; humans set goals, made tradeoffs, and approved every change
- Focus on four journeys covering 95% of user activity
- 3,000+ merged changes with zero customer-facing incidents or rollbacks
The post shares prompts, measurement methods, and the safety loop they built — a strong case study in running agent-driven production performance work.
More from coding & agent
- Google paper: auto-optimized agent harnesses overfit evals — RRSI fixes it — omarsar0 · 2026-09-24
- New video series on agent evals kicks off with episode 1: how to read traces — doesdatmaksense · 2026-09-24
- Formally verifying the Claude Agent SDK with Lean yields 16 bug-fix PRs from a few prompts — dyn___ · 2026-09-24
- Open-source agent lead-enrichment stack Jev+Treg undercuts Clay at $0.029 per verified lead — cneuralnetwork · 2026-09-24
- uv author: metadata-free lockfiles still detect stale resolutions — charliermarsh · 2026-09-24
- Running Claude vs Codex debates in a shared doc works, but it's slow and expensive — kshitizsriv · 2026-09-24