
Agentic AI
📱 Google Makes Gemini a Task-Executing Assistant
What happened
Google announced that Tasks in Gemini is rolling out for free across the U.S., Japan, Korea, Canada, Australia, and India, powered by its Gemini 4 Hydrogen model. Gemini can call businesses, schedule appointments, analyze bills, and shop through more than 60 connected apps, with payment support using the Agent Payments Protocol (AP2).
Why it matters
Google is moving consumer AI beyond answering questions into completing real-world tasks across applications. Hardware-isolated execution environments, configurable permissions, and user approval for sensitive actions are becoming essential as agents gain access to personal information and payment systems.
What's next
Google plans to integrate the Universal Commerce Protocol to improve agent-driven shopping and checkout. Advanced capabilities are beginning in the U.S., while broader availability will depend on the ongoing rollout and integration with external services.
🛡️ Microsoft Deploys Red, Blue, and Green Security Agents
What happened
At Microsoft Ignite, Microsoft announced general availability of Perception, an agentic security system combining red-team agents that simulate attacks, blue-team agents that detect threats, and green-team agents that remediate issues. Its playbooks coordinate these agents across security workflows, while expanded Agent 365 capabilities add cross-platform identity, risk management, and business analytics.
Why it matters
Microsoft is pushing cybersecurity toward coordinated AI-versus-AI operations rather than isolated detection and response tools. Connecting offensive simulation, defensive detection, and remediation creates the potential for continuous security validation, but also raises the stakes for independent testing and human oversight.
What's next
Microsoft is integrating its new Cyber 2 security models with Perception and expanding support for third-party agents through Entra Agent ID. Its newly announced AI Transparency Initiative also aims to establish industry standards for independently verifying AI behavior and controlling autonomous actions.
🔐 Proton Gives Private AI Agents Permission Controls
What happened
Proton released Lumo Desktop for Windows, macOS, and Linux, introducing Connectors, Schedules, and reusable Skills. The open-source desktop assistant connects with local files, Proton Mail, Google Drive, Jira, Confluence, and MCP servers, while allowing users to specify which actions run automatically, require approval, or remain blocked.
Why it matters
As assistants gain access to sensitive files and applications, privacy and permission enforcement become product differentiators. Proton is positioning encrypted, user-controlled AI as an alternative to assistants that rely on broad access to cloud-hosted personal and enterprise data.
What's next
Lumo Desktop is available for download, with connectors and recurring tasks configurable through its interface. Proton is also bringing a new NVIDIA B300 compute cluster online in Switzerland to strengthen its control over the infrastructure behind its AI services.
Generative & Enterprise AI
🚀 OpenAI Releases GPT-6.2 Astra and Sol
What happened
OpenAI introduced GPT-6.2 Astra and GPT-6.2 Sol, upgrading professional reasoning, coding, computer use, and document generation. On OpenAI's Agents' Last Exam benchmark, Astra scored 5.2 points higher than GPT-6 Astra at less than one-third of the cost per task, while Sol approaches Astra's coding performance at substantially lower cost.
Why it matters
The economics of frontier intelligence are changing as models deliver stronger results with fewer tokens and less computation per completed task. This makes long-running agentic workflows, professional deliverables, and complex software development more financially practical at enterprise scale.
What's next
Both models are available through ChatGPT Work, Codex, and the OpenAI API for eligible users. GPT-6.2 Astra's standard API pricing starts at $8 per million input tokens and $40 per million output tokens, with additional savings for cached inputs.
💻 Microsoft Launches Specialized Coding and Cyber Models
What happened
Microsoft introduced Code 2 and Code 2 Flash, specialized AI models for software development, alongside Cyber 2 Flash for cybersecurity work. Code 2 targets complex debugging and multi-file implementation, while Code 2 Flash emphasizes high-volume, lower-cost tasks; both coding models are generally available through GitHub Copilot and Microsoft Foundry.
Why it matters
Microsoft is challenging the assumption that a single frontier model should handle every enterprise workload. Specialized models tuned for coding and security may deliver better task economics, allowing enterprises to reserve more expensive general-purpose intelligence for work that truly requires it.
What's next
Microsoft is integrating Cyber 2 with its Perception security system and introducing MAI Licensed Weights for trusted customers through its Frontier program. The broader strategy is to match specialized intelligence to specific enterprise tasks while reducing the cost of execution.
⚖️ OpenAI Expands Astra for Law
What happened
OpenAI released an updated Astra for Law, combining GPT-6.2 Astra with specialized legal research instructions and an expanded legal search index. On Vals AI's Legal Research Bench, the company reports that answers receiving full credit increased from approximately 46% in September to 53.8% in its November 16 evaluation; all ChatGPT Enterprise customers can now request access.
Why it matters
Enterprise AI increasingly depends on domain-specific context, evaluation, and workflows layered on top of powerful foundation models. The improvement suggests that specialized search and task-specific configurations can deliver meaningful accuracy gains without developing an entirely separate foundation model.
What's next
Organizations can request Astra for Law through the Trusted Access Program, including API access for custom applications. Wider deployment will test how effectively these research improvements translate into legal workflows that still require professional review.
Physical AI
🦿 Boston Dynamics Brings Atlas Into Semiconductor Manufacturing
What happened
Boston Dynamics and GlobalFoundries expanded their robotics partnership, deploying Spot inspection robots across semiconductor facilities in the U.S., Europe, and Asia. GlobalFoundries will also become the first semiconductor manufacturer to participate in the Atlas humanoid early-adopter program; its two Spot robots in Singapore have already completed more than 90,000 inspections with a reported 99% mission success rate.
Why it matters
Physical AI is moving beyond demonstrations into demanding industrial environments where reliability, safety, and operational continuity matter. GlobalFoundries provides a particularly revealing test because semiconductor manufacturing requires precision, repeatability, and minimal downtime.
What's next
Atlas testing is scheduled to begin at GlobalFoundries facilities in early 2027, with workshops assessing humanoid use cases and integration requirements. Meanwhile, the companies plan to expand Spot's inspection responsibilities across the global manufacturing network.
📦 ROBOTERA Scales Humanoid Sorting Across China Post
What happened
ROBOTERA announced that its humanoid robots are operating across 11 China Post postal hubs in eight cities in Guangdong Province. The company reports that robots participating in its Guangzhou pilot, which began in March, achieved an average sorting induction rate exceeding 1,500 parcels per hour, comparable to experienced human workers.
Why it matters
Humanoid robotics is beginning to move from controlled pilots to geographically distributed operations. Logistics provides a measurable test of embodied intelligence because throughput, uptime, handling accuracy, and labor requirements directly determine economic viability.
What's next
ROBOTERA is extending its logistics deployments beyond Guangdong, adapting systems to different parcel types, operating conditions, and sorting-line configurations. The next challenge is demonstrating consistent performance and economics across an increasingly diverse industrial fleet.
💡 The Bottom Line
AI's competitive advantage is moving from intelligence itself to the ability to deploy it reliably at scale. Google and Microsoft are expanding autonomous execution while specialized models reduce the cost of complex work. In physical AI, industrial deployments are turning reliability and productivity into the benchmarks that matter. The emerging moat is not simply smarter AI, but trusted systems that can act, adapt, and deliver measurable results.
⚙️ Try It Yourself
Give Gemini a Task, Not a Question
Google's new Tasks in Gemini demonstrates the difference between conversational AI and delegated work. Try it through Gemini with a Google account; the free feature is rolling out across supported countries, so it may not appear for everyone immediately.
Try this: Open Gemini and switch from Chat to Task. Paste the following:
*****
Find three free, publicly accessible AI courses suitable for someone learning about AI agents. Compare the provider, estimated time commitment, prerequisites, and direct enrollment links. Recommend one for a business professional. Do not register, contact anyone, or make purchases. Ask for my approval before taking any action beyond research.
*****
Look for: A completed comparison with verifiable links, rather than a generic answer. Check whether the agent respects your instruction not to enroll or contact anyone.
Insight: An agent's value is not just finding information, but completing a bounded task while respecting where its authority ends. This simple exercise illustrates delegation, not proof of technical permission enforcement.
