HeadsUpAI

Anthropic Publishes AI Agent Evaluation Framework from Production Deployments

· Updated

Anthropic published a guide to evaluating AI agents covering coding, conversational, research, and computer use categories. It defines three grader types - code-based (test suites, outcome verification), model-based (rubric scoring, LLM judges), and human review - plus the distinction between capability evals and regression evals. Real benchmarks like SWE-bench Verified (where LLMs jumped from 40% to 80%+ in one year) illustrate evaluation in practice.

The guide draws from Claude Code development and collaborations with Descript, Bolt, Stripe, and Shopify. Descript built evals around three dimensions: don't break things, do what was asked, do it well. Bolt built their system in three months using static analysis, browser agents, and LLM judges.

Five frameworks reviewed: Harbor, Promptfoo (used internally at Anthropic), Braintrust, LangSmith, and Langfuse. Start by sourcing tasks from real agent failures.

View the full update on anthropic.com
Anthropic
Anthropic
@AnthropicAI
X

New on the Anthropic Engineering Blog: Demystifying evals for AI agents. The capabilities that make agents useful also make them more difficult to evaluate. Here are evaluation strategies that have worked across real-world deployments. https://t.co/UD0yGglTU0

382retweets
View on X

Share this update