Anthropic: Claude Code can now build evaluations and hillclimb on them to improve applications

A Medium post introduces core/agent-eval, a Claude Code package the authors built to evaluate LLM agent workflows offline before production. The package reads an agent project, auto-detects what matters, and generates an evaluation suite.

Evals Coach is a Claude plugin (skill) that guides product managers through writing good evaluations for LLM features. It uses Claude to coach users on defining test cases and criteria for AI products. It addresses the gap that PMs often lack expertise in designing evals.

Social·Community

Automating EDD (eval-driven development) with Coding Harness Agents

A video tutorial teaches eval-driven development (EDD) for LLM apps: build a baseline application, score every change with LLM-as-a-judge evals, and keep improvements while rejecting regressions. The approach builds on Confident AI's EDD article and Eugene Yan's product-evals write-up, with source code and Agent Skills open-sourced.

A video tutorial teaches eval-driven development (EDD) for LLM apps: build a baseline application, score every change with LLM-as-a-judge evals, and keep improvements while rejecting regressions. The approach builds on Confident AI's EDD article and Eugene Yan's product-evals write-up, with source code and Agent Skills open-sourced.

Anthropic published an engineering writeup on making claude.ai 3x faster within two weeks. The team used Claude to measure, debug, and improve performance, and shared the prompts and methods. The post offers practitioners a reproducible playbook for LLM-assisted performance engineering.

Anthropic is reducing Claude Code usage limits by 50% on Aug 31, prompting a user to consider switching to OpenAI's Codex. The user currently relies on Claude Code's Max 5x plan for orchestration and building, which they say will become unusable after the cut.