
Agentic AI
🏥 OpenAI Brings Computer-Use Agents Into Healthcare
What happened
OpenAI expanded ChatGPT for Healthcare with computer-use agents capable of navigating electronic health records, insurer portals, and other existing software. The release also includes improved clinical search and more than 20 new healthcare plugins, with Abridge, Ambience, Cigna Healthcare, and CVS Health among organizations deploying agent workflows.
Why it matters
Healthcare AI is moving beyond answering questions to executing administrative and clinical-support tasks across disconnected systems. That shift makes application permissions, audit trails, and human approval critical to safely scaling autonomy.
What’s next
Computer-use agents are available through ChatGPT Work and Codex for eligible healthcare users, subject to workspace settings. Healthcare organizations will need to configure access controls, clinical review, and privacy safeguards before expanding deployments.
🧩 Genpact Introduces Earned Autonomy for Enterprise Agents
What happened
Genpact announced the availability of Genpact Onyx, an enterprise platform for building, running, and governing AI agents. Powered by its Process Intelligence Core, the platform embeds operational knowledge into agents and introduces earned autonomy, allowing agents to progress from human-approved recommendations toward greater decision-making authority.
Why it matters
The enterprise AI advantage is shifting from deploying more agents to encoding the processes, exceptions, and business rules that make them reliable. Governance becomes part of how autonomy is granted, rather than a separate compliance layer.
What’s next
Genpact is starting with its Accounts Payable Suite while supporting custom client deployments across existing technology stacks. The longer-term test will be whether reviewed exceptions and human decisions consistently improve agent performance.
🛡️ Corma Raises $125 Million for Autonomous Cyber Defense
What happened
Corma announced a $125 million Series A at a $1 billion valuation, just three months after emerging from stealth. The company develops foundation models specialized for defensive cybersecurity, powering agents that work alongside security teams and execute tasks using their existing tools.
Why it matters
Cybersecurity is becoming a proving ground for specialized agentic AI, with defensive models designed to respond to increasingly automated threats. Corma's rapid financing underscores investor conviction that security teams will need AI systems capable of executing work, not simply analyzing alerts.
What’s next
Corma plans to use the funding to expand compute capacity, scale its foundation models, and selectively grow its team. The company says another generation of models is coming, although it has not announced a release date.
Generative & Enterprise AI
🚀 Alibaba Launches Qwen4-Max With Lower Operating Costs
What happened
Alibaba released Qwen4-Max, its new flagship model featuring 2.4 trillion total parameters, 71 billion activated parameters per token, native visual understanding, and a one-million-token context window. On its proprietary Work360 benchmark, Alibaba reports stronger performance than Qwen3.8-Max while using 35% fewer tokens at high reasoning effort and reducing average task costs by approximately 60%.
Why it matters
AI competition is increasingly about the cost of completing useful work, not simply the size or intelligence of a model. More efficient reasoning and stronger agentic capabilities could make complex, long-running enterprise workflows economically practical at greater scale.
What’s next
Qwen4-Max is available through QwenCloud, QwenWork, and QwenStudio. Alibaba also plans to release smaller open-weight Qwen4 models, potentially extending the architecture to developers seeking more deployment flexibility.
🔬 NVIDIA Gives Scientific Agents Specialized Intelligence
What happened
NVIDIA introduced Nemotron for Science, combining its Nemotron 3.5 Super VL open model with specialized training recipes, scientific evaluations, and LitmusBench v0.2, a chemistry dataset containing nearly 50,000 questions. Research organizations including Argonne National Laboratory are integrating Nemotron into agents for scientific discovery.
Why it matters
The next wave of AI capability may come from adapting powerful foundation models to specialized domains rather than relying entirely on larger general-purpose systems. Scientific agents that combine domain knowledge, evaluation, and computational tools could accelerate research while reducing dependence on expensive frontier models.
What’s next
The model and chemistry dataset are available on Hugging Face, allowing researchers to develop and evaluate specialized scientific agents. NVIDIA is also expanding integrations with research institutions and laboratory workflows, where independently validated results will determine their practical value.
Physical AI
🦾 Siemens Turns Engineering Prompts Into Robotic Grippers
What happened
Siemens announced an agentic engineering workflow developed with AGIBOT that transforms natural-language requirements into customized robotic gripper designs. AI agents orchestrate geometry adaptation, simulation, interface validation, and manufacturing preparation through Siemens engineering software and Model Context Protocol (MCP) integrations.
Why it matters
Industrial AI is beginning to connect human intent directly to engineering execution. By coordinating multiple design tools and preparing production-ready data, agents could reduce the manual effort required to customize physical components for different applications.
What’s next
Siemens plans to demonstrate the robotic gripper workflow at Formnext 2026 in Frankfurt. The company is also developing an adaptive additive-manufacturing production cell with EOS and AM Solutions to connect engineering, manufacturing, and automated post-processing.
🏭 Aptiv Targets 100 Humanoid Robots on Factory Floors
What happened
TARSBOT and Aptiv detailed an expanded embodied-AI partnership targeting deployment of 100 industrial robots by the end of 2026 across eight manufacturing process categories and 33 application scenarios. According to the companies, the first TARSBOT A-Series robots are already operating in an Aptiv factory.
Why it matters
Humanoid robotics is entering a more consequential phase where deployment scale, uptime, and operational productivity matter more than demonstrations. Real factory environments also create feedback loops that can improve robot perception, decision-making, and task execution.
What’s next
The partners plan to expand validation across additional production workflows while integrating Aptiv computing and power-management technologies into TARSBOT's robots. Their broader collaboration targets a 1,000-unit industrial robot fleet, although reliable performance at that scale remains to be demonstrated.
💡 The Bottom Line
AI's competitive advantage is shifting from generating intelligence to executing work reliably. OpenAI and Genpact are embedding agents into governed workflows, while Alibaba and NVIDIA are pushing efficiency and domain specialization. In physical AI, Siemens and Aptiv are demonstrating how that intelligence connects to real production, where measurable outcomes, not model benchmarks, will ultimately determine adoption.
⚙️ Try It Yourself
Turn a Prompt Into an Operational Workflow
Alibaba's Qwen4-Max release highlights AI's growing ability to handle complex professional work. Experience the shift using QwenWork, which offers a browser-based free plan with signup credits.
Try this: Create a free account, open a new task, and paste this prompt:
*****
Create a one-page operating procedure for a fictional online retailer handling customer refunds. Include five workflow steps, task owners, exception handling, and approval checkpoints. Identify which steps an AI agent could perform independently and which require human authorization. Use a table and make reasonable assumptions explicit. Do not access real accounts or process transactions.
*****
Look for: A structured document with clear responsibilities, escalation rules, and approval gates. This is a planning demonstration, not proof that permissions would be enforced in a live system.
Insight: Building an AI workflow is increasingly easy. Defining where autonomy begins and ends is the harder business problem.
