Claude Science, AI agents as coworkers, and the data problem no one wants to talk about

Three stories this week that cut through the noise: a major product launch aimed at research-intensive industries, a necessary correction to how executives are framing AI agents internally, and a sector-specific warning about data readiness that applies far beyond farming. Each one has something direct to say about how companies should be making decisions right now.

Anthropic launches Claude Science and aims directly at pharma and biotech

At a closed event for pharmaceutical executives, biotech founders, and researchers, Anthropic unveiled Claude Science, a product built specifically to support scientific research workflows. The positioning mirrors what Claude Code did for software development: a dedicated, domain-aware tool rather than a general-purpose assistant. Early focus is on industries where research cycles are long, regulatory stakes are high, and the cost of a wrong answer is measured in years and dollars, not just inconvenience. For executives in life sciences, diagnostics, or any research-heavy operation, this signals that the most credible AI labs are moving from horizontal platforms toward vertical depth. If your industry has not yet seen a purpose-built AI product aimed at your core workflow, it is coming. The question is whether you will be ready to evaluate it when it arrives. MIT Technology Review

AI agents are not your coworkers, and managing them like they are will cost you

MIT Technology Review published a piece this week arguing that the "AI as coworker" framing that has spread through executive communications and internal memos is doing real damage. The argument is straightforward: treating agents as junior employees creates wrong expectations about reliability, accountability, and oversight. Agents do not push back, do not flag ambiguity unless explicitly designed to, and do not carry context the way a person does across weeks of work. When something goes wrong, there is no conversation to have. The organizational risk is that companies design handoff points and review processes around a metaphor rather than around how these systems actually behave. The practical fix is to define what each agent does, what it cannot do, and who owns the output, before it touches anything that matters. MIT Technology Review

Agriculture's AI moment is being blocked by its own data infrastructure

A detailed MIT Technology Review analysis makes the case that agriculture is one of the sectors with the most concrete AI use cases, from yield prediction to pest detection to supply chain optimization, but that most operations cannot yet benefit because their data is fragmented, inconsistently labeled, or simply not collected in a usable form. The piece is written about farming, but the diagnosis maps cleanly onto manufacturing, logistics, healthcare administration, and any industry where operational data has historically lived in spreadsheets, paper records, or siloed systems. The pattern is the same everywhere: companies try to deploy AI on top of messy data and then conclude that AI does not work for their context. The real conclusion should be that data readiness is not a prerequisite someone else handles. It is a strategic project that needs a budget, an owner, and a timeline before the AI conversation begins. MIT Technology Review

Google's NotebookLM adds short-form video output, and that changes the use case

Google is rolling out a feature in NotebookLM that converts research notes and documents into 60-second vertical video clips, available initially to AI Ultra and Pro subscribers. The format follows TikTok conventions. What is worth paying attention to here is not the format itself but the workflow change: NotebookLM was already useful for turning long documents into audio summaries. Adding a visual output layer means teams can now go from internal report to shareable briefing clip without a production step. For marketing, communications, and knowledge management teams, this reduces the friction between insight and distribution. The limitation is that it is still AI interpreting your source material, so the same content review discipline that applies to generated text applies here. The Verge

Tidal draws a line on AI music: label it, but do not pay for it

Starting July 15th, Tidal will identify AI-generated tracks on its platform and strip them of royalty eligibility, while stopping short of removing them entirely. The policy is notable because it is one of the first major platform decisions that treats AI-generated content as a distinct category with distinct commercial rules, without banning it outright. For companies building content operations that involve AI-generated audio, including branded podcasts, background music for video, or audio ads, this is a preview of how distribution platforms will handle the question. The pattern will likely repeat: tolerate the content, label it, deny the economics. If your content strategy depends on AI-generated audio reaching monetized distribution channels, that window is narrowing. The Verge

What to take from this week: Anthropic is making a serious vertical bet on science, which is a signal about where specialized AI value will concentrate. The agent-as-coworker framing is a liability, not a communication strategy. And data quality is still the unglamorous bottleneck that determines whether AI projects succeed or get quietly shelved. None of these are new problems, but this week gave each of them a sharper edge.