Claude Goes Background, Agents Go BYO Infra

Claude can work your desktop in the background

Claude Cowork and Claude Code can now click, type, and open desktop apps while you work on something else.

• The feature is available in beta for Pro and Max plans in the Claude desktop app on macOS.
• You can enable computer use in Settings → General → Computer use.
• The pitch is simple: hand Claude a desktop task, then keep moving on your own work.

What it means Start testing Claude on repeatable desktop chores, but define tight task boundaries and review points before you let an agent touch work that matters.

Source: claudeai

Cursor agents can run on your own infrastructure

Cursor cloud agents can now run on infrastructure you manage, including machine pools that scale with demand.

• Agents can reach internal services or specialized hardware while Cursor keeps the agent loop.
• You can run them on your own machines or through supported sandbox providers.
• Supported providers include AWS Lambda, Coder, Cloudflare, Daytona, E2B, Modal, Namespace, and Vercel.

What it means Move agents closer to your internal tools and data, then use managed machine pools to test workloads that need private services or custom hardware.

Source: cursor_ai

Gemini 3.8 Flash targets multi-step agent work

Google introduced Gemini 3.8 Flash, calling it its most intelligent workhorse model for software engineering, agentic tasks, and multi-step reasoning.

• Google AI Studio says it improves on Gemini 3.7 Flash across software engineering and specialized multi-step tasks.
• It is available at the same introductory price as Gemini 3.7 Flash.
• Google says builders can use it through the Gemini API via Google AI Studio and Android Studio.

What it means Run your existing Flash prompts through 3.8 now, especially coding and multi-step flows, because Google says you can test the upgrade without a price increase.

Source: GoogleAIStudio

Perplexity open-sources its local Mac inference engine

Perplexity has open-sourced Lily, the local inference engine it built for hybrid compute in Perplexity Computer.

• Lily is specialized for Qwen3.6-35B-A3B on Apple Silicon.
• Hybrid compute splits work between cloud models and a local Mac model.
• Perplexity says Lily averaged 1.23× higher prefill throughput and 1.35× higher decode throughput than MLX-LM in its M5 Max MacBook Pro benchmark, with output quality effectively unchanged.

What it means If you build Mac-first AI products, evaluate Lily for Qwen3.6-35B-A3B workloads where local inference needs to keep up with cloud-assisted tasks.

Source: perplexity_ai

Replit turns analytics on for new projects

Replit now enables project analytics for every new project, giving builders a clearer view of how people engage with published apps.

• Existing projects can enable analytics through Publishing → Adjust Settings → Enable Analytics.
• You can ask Replit Agent to suggest custom metrics for your project.
• Replit frames the data as a way to guide the next feature you build.

What it means Add analytics and a few custom events before your next launch so user behavior, not just feedback, tells you what to fix or build next.

Source: Replit

Is Fable 5.1 About to Change Everything? (Deep Dive)

A detailed look at Fable 5.1 and what it could mean.