AI agents are moving into the browser and the team chat

Google integrated computer use into Gemini 3.5 Flash, letting developers build agents that inspect interfaces and act across browsers, desktops and mobile environments. Anthropic launched Claude Tag in beta, bringing a team agent into Slack to remember channel context and carry out delegated tasks.

Google offers optional confirmation for sensitive actions; Claude Tag lets administrators scope tools and information to selected channels.

Top Stories#

OpenAI unveiled its first custom chip for running AI models. OpenAI and Broadcom unveiled Jalapeño, OpenAI’s first custom chip for running AI models, on June 24. Engineering samples are running models in the lab, including GPT-5.3-Codex-Spark. Initial deployment is planned for the end of 2026.

The design focuses on moving data between computation, memory and networking. OpenAI says the team went from initial design to manufacturing tape-out in nine months, the milestone at which a chip design is sent for fabrication.

Final performance measurements are still underway, with a technical report promised in the coming months. OpenAI describes Jalapeño as the first in a series of chips developed with Broadcom; Celestica is handling board, rack and system integration.

Google put computer interaction inside Gemini 3.5 Flash. Developers can use the integrated capability to build agents that inspect interfaces and act across browsers, desktops, and mobile environments, without a separate computer-use model. Google names continuous software testing and long-running work across applications as uses.

Access comes through the Gemini API and its enterprise agent platform. Optional safeguards request confirmation for sensitive actions and stop tasks when indirect prompt injection is detected; Google still recommends sandboxing, access controls, and human verification.

Claude Tag brings delegated tasks into Slack. Anthropic launched the team agent in beta for Team and Enterprise plans. Claude Tag remembers channel context, uses authorized tools and codebases, and completes work asynchronously.

The announcement describes shared conversations, optional proactive updates, and tasks that continue over hours or days. Administrators can scope tools, information, and memories to selected channels, set spending limits, and inspect activity logs.

Mistral’s OCR 4 returns document structure alongside extracted content. The release adds text locations, document-block types, and confidence scores, with support for 170 languages and self-hosting in a single container. The added structure supports highlighting, citations, redaction, invoice processing, and human review.

The API costs $4 per 1,000 pages, or $2 with batch processing. Self-managed deployment is available to enterprise customers.

Meta paused employee tracking after an internal data exposure. WIRED reports that information collected to train AI became broadly accessible internally. The program gathered clicks, keystrokes, mouse movements, and screen content.

Meta confirmed the pause while investigating and said it had no indication employees improperly accessed the data.

Top Links#

• Unlimited OCR Works (arxiv.org): The authors propose a constant-size attention cache to address memory growth during longer transcriptions, while retaining a 32K output limit.
• Patch the Planet (openai.com): OpenAI and Trail of Bits combine AI-assisted vulnerability discovery with human review and patch development. OpenAI reports work across 19 projects and dozens of merged patches.
• European provider diversification (kuwaittimes.com): Reuters reports that US service restrictions are reinforcing European companies’ efforts to retain multiple providers and locally runnable alternatives.
• Cooling servers with warm liquid (blogs.nvidia.com): NVIDIA explains how coolant as warm as 45°C can release heat through outdoor radiators, reducing dependence on mechanical chillers and evaporative cooling.