Agentic AI systems are proving both powerful and fragile: a new coding agent shows strong promise for low-cost automation, but new research reveals LLM agents systematically fail to maintain constraints during backend code generation — a critical risk for regulated environments
Microsoft's admission that AI agents can cost more than human employees reframes the ROI conversation — AI PMs must build rigorous cost-tracking into agent deployments before scaling, not after
AI governance is moving into new institutional territory: Anthropic's co-founder appearing alongside Pope Leo XIV at a Vatican encyclical on AI signals that ethics and regulation are converging at the highest cultural levels, which will shape enterprise compliance expectations
Microsoft has publicly acknowledged that token costs and agent overhead in some workflows exceed the cost of equivalent human labour, surfacing a fundamental tension in the business case for AI automation
An open-source project wraps DeepSeek's reasoning model into a coding-agent workflow with aggressive prompt caching designed to cut inference spend substantially, attracting significant community traction (381 HN points)
arXiv paper 2605.06445 finds that LLM agents progressively lose track of user-specified constraints as code-generation tasks grow longer or more complex — a failure mode dubbed "constraint decay."
Anthropic co-founder Chris Olah has been invited to stand alongside Pope Leo XIV at the release of a papal encyclical on AI (due May 27), the result of ongoing Anthropic–Vatican dialogue on AI ethics
Willison fed a scanned Usborne "Creepy Computer Games" PDF to Claude and had it produce a playable version of "Mad House," illustrating multimodal document-to-code pipelines in a low-stakes, creative context