📈 AI Storyline

TL;DR

Deep research agents — systems that run multi-step research using dynamic reasoning and multi-hop retrieval — moved from field notes toward repeatable practice over three weeks. InfoQ opened with Sarang Kulkarni's production lessons on building them; an arXiv paper then flagged an evaluation gap, arguing existing benchmarks only score single-shot output and proposing multi-turn evaluation under process-level feedback; AWS closed with an end-to-end build using Deep Agents and Bedrock AgentCore.

What's new: The latest beat is AWS's hands-on walkthrough building a competitive research agent end to end on Deep Agents plus Bedrock AgentCore — the thread's first concrete, platform-specific build pattern, after the earlier production-lessons and evaluation-method posts.

💡 Deep research agents are maturing into something an engineer can actually ship: there's now a managed-cloud build pattern to copy and a multi-turn evaluation method to judge whether the agent improves under feedback rather than just its one-shot answer.