Anthropic Expands Claude Cowork to Mobile and Web — Tasks Keep Running Even When You Close Your Laptop

On July 7, Anthropic announced that it is extending its AI agent, Claude Cowork, to mobile and the web. Users can now start a task on desktop, check its progress from their smartphone, and receive the finished output later. Even with the laptop closed, the agent keeps processing in the background.

TechCrunch described the move as "the coding agent war expanding to the entire office," and WIRED covered it as part of a broader trend of "controlling agents from your smartphone." What stands out is that agent use cases are expanding from developers' desks to every corner of daily life.

Microsoft Starts Swapping OpenAI/Anthropic Models for Its Own AI in Excel and Outlook

According to a July 7 Bloomberg report, Microsoft has begun replacing some OpenAI and Anthropic models in software products such as Excel and Outlook with its in-house "MAI" models in an effort to cut AI costs. Tens of thousands of AI prompts are reportedly already being processed by MAI models every week.

The company has invested heavily in OpenAI for years, but with rising AI infrastructure costs, it is pushing ahead with a shift toward models it built itself. The replacement appears to be limited to certain features for now, but it signals that major tech companies are accelerating efforts to rethink their AI cost structures.

Google Massively Expands Managed Agents in the Gemini API

On July 7, Google announced new capabilities for Managed Agents in the Gemini API. The key additions include:

Background tasks**: Run long-running processes asynchronously

Remote MCP support**: Integrate external tools via the Model Context Protocol

Additional reliability improvements aimed at production use

The building blocks developers need to create "production-grade agents" are coming together. Google, Anthropic, and AWS are all pouring effort into strengthening their agent-building platforms, and competition in this space is heating up in earnest.

Liquid AI Releases "Antidoom" — Eliminating Doom Loops in Reasoning Models

On July 7, Liquid AI released "Antidoom," an open-source method for eliminating the "doom loop," a common failure mode in reasoning models. A doom loop is the phenomenon where a model repeatedly cycles through the same reasoning pattern and its output stalls.

According to the published benchmarks, doom loop occurrence rates dropped sharply:

LFM2.5-2.6B (early checkpoint)**: 10.2% → 1.4%

Qwen3.5-4B**: 22.9% → 1% (with greedy sampling)

While doom loop rates fell, evaluation scores improved overall — a significant result that should make reasoning models more practical.

mistral.rs v0.9.0 — Up to 1.8x Faster CPU Inference Than llama.cpp

Version 0.9.0 of mistral.rs, an LLM inference engine written in Rust, has been released. It reportedly delivers CPU decoding up to 1.8x faster than llama.cpp on both x86 and ARM architectures.

As local LLM adoption grows, inference performance on CPUs directly affects real-world practicality. More options for running models quickly in GPU-constrained environments is good news for edge AI and privacy-sensitive use cases.

Data Center Construction Delays Threaten Global AI Growth

On July 7, The Guardian reported that data center construction projects around the world are falling behind, threatening the growth of the AI industry. Power supply constraints, environmental regulations, and rising construction costs are cited as the main factors.

Compute demand is soaring as AI models scale up and agent usage grows, but the buildout of supporting infrastructure isn't keeping pace. As a structural challenge bearing on the sustainability of the AI industry, its medium- and long-term impact will be worth watching.

---