How Anthropic's team ran a two-week performance sprint with Claude in Slack, and how to copy the loop. Source: Anthropic's post "How we made claude.ai 3x faster in two weeks" (claude.dev, September 23, 2026).
The results
| Journey (p75) | Before | After |
|---|---|---|
| Fresh load of claude.ai | 3.1 s | 0.55 s |
| Desktop app cold start | 6,310 ms | 3,328 ms |
| Loading a conversation (web) | 1,557 ms | 646 ms |
3.1x faster on average across thirteen measurements. More than three thousand changes merged, with no customer-facing incident or rollback.
The loop
- Give the agent a standing job. Their channel told Claude to monitor deploys for regressions, keep dashboards curated, fix low-hanging fruit, propose projects and talk to human teammates.
- Measure first. "As soon as Claude had a number to beat, it could start optimizing." Prefer deterministic counts over noisy timings: React commits per interaction, function call counts, layout and style-recalc counts, DOM mutations.
- Distrust every benchmark. If a benchmark is flaky or doesn't track real user latency, throw it out so the agent doesn't "climb the wrong hill."
- Keep threads narrow. One benchmark or journey per thread, with a named human owner.
- Show, don't tell. The agent attaches before-and-after screenshots or recordings for any change a user could notice, and the owner rules on it.
The guardrails
- Unit tests come before optimizations.
- Every PR gets automated review plus at least one human approval.
- Risky changes ship behind a short-lived feature flag, labelled as a kill switch or a ramp, and retired once safe.
- Roll out in stages: employees, then one percent of users, then everyone.
- After it ships, the agent watches the deploy and reads the field data.
Protect the win
Turn each benchmark into a CI guardrail whose number "could only ratchet down," so later changes can't quietly undo the gain.
Starter instruction (adapted from Anthropic's)
Your job is to own the performance of [app]. Monitor deploys for regressions, keep our dashboards accurate, fix low-hanging fruit, propose bigger projects, and check in with the thread owner before any user-visible change.
Humans still set the goals and make the taste calls. The agent does the climbing.