claude-code

Anthropic's research team spent a year building evals for Claude Code and multi-agent systems. Here's what actually works: start with 20-50 real failures, combine deterministic and model-based graders, and read the transcripts. Teams that invest early ship faster.

Anthropic's performance engineering take-home has been redesigned three times because each new Claude model solved it. The latest version uses deliberately weird puzzles to stay ahead of AI — because realism is now a luxury they can't afford.

Anthropic researcher ran 16 parallel Claude instances for 2 weeks to build a 100,000-line C compiler that boots Linux. No human intervention. $20k in API costs. This is what autonomous agents can barely do today — and what they'll do routinely tomorrow.

Claude 3.5 Sonnet hit 49% on SWE-bench Verified with a minimal agent scaffold: two tools, one prompt, maximum model control. Here's the exact architecture Anthropic used and why tool design matters as much as model capability.

Anthropic's "think" tool gives Claude structured reasoning checkpoints during execution. It delivered 54% improvement on policy-heavy tasks and 1.6% on SWE-Bench. Here's when to use it—and when to skip it.

Anthropic's Claude Code runs in terminal, IDE, browser, and mobile with session portability across all surfaces. Same context, different environments.