Agent Tooling Gets Real, Loop Architectures Evade Benchmarks, and The Memory Tax Hits Production

Documentation

Anthropic Updates Claude Tool Use Documentation with Production Patterns

Anthropic's updated docs now explicitly cover error handling, retries, and streaming for tool use. This is them admitting that the first version was for demos, and this version is for people who actually have to wake up at 3 AM when an agent loops.

Analysis

BAIR: Why Compound AI Systems Will Outperform Single Models

This is the paper everyone cites but nobody reads. The core argument is that orchestrating multiple models and tools together will always beat a single monolithic model, and the research backing is solid. If you're still trying to make one model do everything, you're fighting physics.

News

The Batch: AI Agent News and Industry Updates

Andrew Ng's latest roundup covers the agent landscape with a focus on practical deployment. Good filter for what's signal vs. hype, especially the take on agent evaluation frameworks.