Cursor trains Composer in production by serving model checkpoints, observing user responses, and using those responses as reward signals to ship improved versions every 5 hours.

USV created internal agents that process meeting transcripts, emails, and calendars to create structured mentions tied to companies and people, forming a continuously updated knowledge base.

Latest Claude Code update adds session ID headers for better proxy handling and fixes resume functionality for interrupted autonomous coding sessions.

Intercom's vertical customer service agent resolves 2 million issues weekly and beats frontier models while generating nearly $100M in recurring revenue.

New benchmark tests interpretability methods on predicting future actions, detecting intervention effects, and identifying distributional properties of agent rollouts.

Agent capabilities are rapidly improving on autonomous benchmarks even as per-task costs increase, raising questions about economic viability of automation.