Windsurf Is Reborn as "Devin Desktop" — Unified Management of Multiple AI Agents

On June 4, Cognition refreshed its AI code editor Windsurf into "Devin Desktop". The biggest change is that agent management capabilities now come built in as standard, letting you centrally manage multiple AI agents — including OpenAI's Codex and Anthropic's Claude Agent — within a single environment.

While the previous Windsurf was limited to integration with a single AI assistant, Devin Desktop reportedly lets you use different agents on a per-project basis and run them in parallel. It also envisions workflows where agents share context with one another and human review is inserted into the loop.

As competition among AI coding tools intensifies, the move suggests that the axis of differentiation is shifting from "which AI to use" to "how to integrate multiple AIs."

VS Code's "Integrated Browser × AI Agent" Is a Paradigm Shift in Development

What is drawing attention on Qiita is the latest VS Code update, which ships an Integrated Browser as a standard feature alongside integration with AI agents such as GitHub Copilot.

With the browser rendered directly inside the editor and AI agents able to "see" and operate that view, development workflows are changing dramatically. The cycle — render code changes in the browser in real time → AI gives visual feedback → apply fixes — can now be completed without hopping back and forth between external tools.

The author describes it as the editor having gained "hands" and "eyes", and productivity gains in frontend development are particularly anticipated.

iPhone's ANE Outperforms MLX for LLM Inference — Thermal Throttling Makes the Difference

Benchmark results published on GitHub are attracting attention. When running an LLM on the iPhone's Apple Neural Engine (ANE), token generation speed is reportedly more stable than with MLX or LiteRT.

While MLX and LiteRT suffer from thermal throttling (performance degradation caused by heat) under heavy load, the ANE's power-efficient design lets it sustain speed even during long-running inference. For use cases that keep LLMs running continuously on edge devices, leveraging the ANE looks set to become a compelling option.

"Phylax" Tackles the Risk of AI Agents Reading and Writing Files

Phylax, a tool that prevents damage caused by autonomous file access by AI agents, is making waves on Hacker News. It works as a security layer that stops AI agents from reading or deleting files outside the user's intent.

As AI coding agents proliferate, the risk of an agent accidentally manipulating critical files has become a real concern. Phylax reportedly enforces policy-based access control and makes agent actions auditable.

API Migration Guide for GPT Image 2 Published

An API implementation guide for GPT Image 2 (gpt-image-2), the successor to DALL-E 3, has been published on Qiita. For this model, officially released in April 2026, the guide carefully explains how to call the image generation API and how to migrate from DALL-E 3.

In addition to the model's performance improvements, it also covers changes to the API interface and new parameters, making it a practical reference for developers working with image generation AI.

The EU's Cloud and AI Development Act — Consideration for Open Source and Remaining Challenges

An analysis of the EU's Cloud and AI Development Act has been published on the SUSE blog. The act aims to strengthen the competitiveness of cloud infrastructure and AI development, but its impact on the open source community is also up for debate.

The article takes a positive view of the act evaluating open source AI from an "Open Source First" perspective, while expressing concern about how the scope of the regulations and compliance costs could affect small-scale developers.

---