November 18, 2026CodingAgentsSkills

Claude Code Team: The Test Suite Is the Real Product

About 90% of the Claude Code team's own code is now written and reviewed by Claude. That is the opening fact of a post the team published on Nov 16, which climbed to 233 points on Hacker News on Nov 17. The rest of the post is about why that number is only possible because of something boring: verification. https://claude.dev/blog/how-the-claude-code-team-thinks-about-verification/

The argument is simple. One agent you can babysit. Dozens or hundreds running in parallel you cannot, so the only thing that scales is a loop that tells the agent, without you, whether it got it right. The team says a separate grader checking output against a rubric lifted task success by up to 10 points on structured file generation. The model did not change. The checker did.

The practical advice is sharper than the usual testing sermon. Flaky tests are worse than no tests, because they teach the agent to rerun instead of fix. Slow tests mean fewer checks per task. End-to-end tests come first because they are hard to fool: give the agent hands and eyes through a browser protocol or a simulator. Property tests, the round-trip, invariant and idempotency kind, catch what example tests miss. And when the same failure keeps coming back, turn it into a hook or a CLAUDE.md rule so the loop closes itself.

The same day, StarSling launched on Hacker News with the other half of the problem: agents now write code faster than CI can check it. It claims Mastra's p95 CI time went from about 30 minutes to about 5 (https://starsling.dev). And Claude Code v2.1.326 shipped on Nov 16 letting Code Mode run with agent teams. Read together, the message is that the bottleneck has moved from writing code to proving it works, and whoever owns the fastest honest checker wins.

This article was written with Claude, an Anthropic model.
← Previous
FrontierHarness v1.1: There Is No Best Harness, Only Best Pairs
Next β†’
PaperOrchestra: Google Open-Sources the Paper-Writing Agent Team
← Back to all articles

Comments

Loading...
>_