AI Stops Watching and Starts Doing

Anthropic has put a new capability into research preview: Claude can now control a Mac directly. Open apps, navigate browsers, scan emails, fill spreadsheets. It works via Claude Cowork and Claude Code, and it's available on Pro and Max plans, macOS only for now.

This is more significant than another model benchmark. AI is crossing from "thing you query" to "thing that operates software." The feature draws on connected apps first - Slack, Calendar - but falls back to direct app interaction when no connector exists.

But, there are real caveats. It's a research preview for a reason. The community reaction was a mix of genuine excitement and pointed concern about fragility. Browser control and computer use in real-world conditions is harder than demos suggest. Practitioners who've pushed these tools into production workflows are already flagging instability.

For construction firms, the immediate thought is probably the admin stack. Chasing submittals, cross-referencing drawing revisions, updating document registers, generating meeting summaries - these are repetitive, time-consuming tasks that sit precisely in the target zone for AI agent control. The capability is real. The reliability, in complex and fast-moving project environments, needs proving.

Quietly, and without much fanfare, document parsing has gone from utility to core infrastructure for AI-powered workflows.

The combination of LlamaParse and Google's Gemini 3.1 Pro is now pulling structured data from difficult financial PDFs with roughly 15% better accuracy than previous approaches - handling tables and complex layouts that would have stumped earlier tools. LlamaIndex has also launched LiteParse, a lighter-weight parsing path designed specifically so that AI agents can call it cheaply and quickly without relying on vision-language models.

This matters for construction because the industry runs on documents. Drawing registers. Specifications. O&M manuals. Contract schedules. RFI logs. The vast majority of this is locked in PDFs, often poorly formatted, often scanned. If AI agents are going to be genuinely useful on construction projects, they need to read these documents reliably.

The accuracy improvements are real. But, the more interesting development is the architectural one - parsing is being designed as something AI agents call as a primitive, not something humans use as a standalone tool. That shift has significant implications for how construction software is built over the next few years.

There's an uncomfortable pattern emerging in the AI practitioner community. Newer, more capable models are sometimes too eager. They delegate tasks to weaker sub-agents, generate parallel outputs that look impressive, and produce what's being called "slop theater" - the appearance of productivity without the substance.

Practitioners running AI in real coding and operational workflows are clear on what separates actual gains from noise: feedback loops. Traces, evaluations, incident tracking, production feedback. Not just generating outputs and moving on.

This is a direct and practical warning for construction firms starting to embed AI into project workflows. It's tempting to measure AI adoption by volume - how many documents got summarised, how many reports got drafted. But, the right measure is accuracy and reliability over time. An AI tool that summarises twenty documents incorrectly is worse than not having the tool at all, especially on a project where those documents feed into programme decisions.

But, the same practitioners making this point are also clear that the technology is maturing fast. The firms that build proper evaluation into their AI workflows now will have a significant advantage over those that chase deployment speed without quality controls.

Two wider AI developments worth noting this week.

Meta's Superintelligence Lab, led by Nat Friedman and Alex Wang, has execuhired the Dreamer team - a licensing-plus-hire deal that keeps the talent and technology without a full acquisition. This follows the $2bn Manus deal in December. Meta is clearly building a formidable consumer AI agent capability. The practical implication for the tools market is that the biggest distribution platform in the world is now seriously invested in agent technology. That tends to accelerate adoption curves.

Separately, data on OpenRouter token usage over the last seven days shows Chinese open-source models dominating - with Xiaomi's MiMo-V2-Pro leading at 1.77 trillion tokens. Cursor (the widely-used AI coding tool) has acknowledged that Kimi K2.5 is currently the strongest open-source base model. Only three Western labs appear in the top rankings. This is a meaningful shift. Open-source models from Chinese labs are now genuinely competitive with frontier models for practical tasks - and they're free to use. For construction technology businesses evaluating AI infrastructure costs, this matters.