Agentic AI

🏢 IBM and Microsoft Package Agents for Regulated Industries

What happened
IBM and Microsoft introduced Imagine Enterprise, a platform for building, running, and governing industry-specific agents. IBM contributes more than 6,000 prebuilt agents and assets, while Microsoft provides the data, identity, security, cloud, and orchestration layers underneath them.

Why it matters
Enterprises increasingly do not need another general-purpose assistant. They need agents that understand existing systems, permissions, business context, and industry requirements without forcing companies to rebuild their technology stack.

What’s next
Imagine Enterprise is available through IBM Consulting or through self-management options, initially targeting financial services, insurance, and retail. IBM and Microsoft say the same framework could expand into manufacturing, energy, government, and other industries.

✅ ServiceNow Turns Compliance Into an Always-On Agent Job

What happened
ServiceNow introduced agentic continuous control monitoring for its Integrated Risk Management platform. Agents can configure checks, interpret failures, open remediation workflows, and automatically fix issues within defined policies across more than 100 integrations spanning cloud, identity, SaaS, containers, on-premises systems, and connected devices.

Why it matters
Compliance traditionally proves that a control worked at a moment in the past. Agentic monitoring turns assurance into a continuous process where the system can detect that a control failed, understand the consequence, and start fixing it immediately.

What’s next
ServiceNow expects general availability in the first quarter of 2027. The bigger test will be whether automated remediation reduces audit work without creating a new layer of autonomous actions that themselves require additional oversight.

🎨 Adobe Lets Agents Work Inside Creative Files

What happened
Adobe expanded its Creative Agent across Firefly and Creative Cloud, allowing users to delegate multi-step creative work in natural language. Adobe also previewed local MCP integrations that will let third-party agents modify native Photoshop, Illustrator, Premiere, and After Effects files while keeping their layers, sequences, and compositions editable.

Why it matters
Agents are moving from generating assets outside professional tools to operating inside the systems where real work is created and maintained. Preserving editable project structure matters because professionals still need to inspect, refine, and control what the agent produces.

What’s next
Adobe’s Firefly AI Assistant and expanded Creative Cloud assistants remain in beta, while local MCP integrations and always-on creative agents are in preview. The next proof point is whether agents can reliably handle recurring creative workflows without sacrificing editability or brand control.

Generative & Enterprise AI

📐 Google Gives Visual Knowledge Work a Harder Benchmark

What happened
Google DeepMind and Artificial Analysis launched the Multimodal Knowledge Agent benchmark, testing agents on 200 projects across 10 industries and 83 occupations where critical information lives inside drawings, diagrams, spreadsheets, presentations, and other visual files. Each task represents work that would typically take an experienced professional three to eight hours.

Why it matters
Professional work rarely arrives as clean text. Agents that cannot interpret page layouts, diagrams, charts, and mixed file types will struggle with engineering, architecture, finance, healthcare, and other workflows regardless of how well they perform on text benchmarks.

What’s next
Artificial Analysis is evaluating 21 models on the benchmark, while Google has released 10 public tasks for outside testing. Current results show substantial differences in quality, cost, token usage, and execution time, reinforcing that model selection increasingly depends on the actual work being performed.

🏥 Google Gives Healthcare Its Own Gemini Agent

What happened
Google Cloud launched Gemini for healthcare, an industry-specific agentic system that connects information from electronic health records, billing platforms, clinical guidelines, and government datasets. The system can reason across that context and execute multi-step clinical and administrative workflows within defined guardrails and human oversight.

Why it matters
Industry AI is becoming less about fine-tuning a general model and more about surrounding it with trusted data, permissions, specialized tools, and workflows. Healthcare raises the bar because errors can affect clinical decisions, payments, privacy, and patient outcomes.

What’s next
Gemini for healthcare is initially available in preview with organizations including HCA Healthcare, Hartford HealthCare, Highmark Health, Optum, and Stanford School of Medicine. Google says patient records and prompts remain inside each organization’s dedicated environment and are not used to train its foundation models.

Physical AI

🤖 Gemini Robotics Moves Onto the Factory Floor

What happened
Google DeepMind introduced Gemini Robotics 2.5, a vision-language-action model evaluated across more than 200 physical tasks. In a proof-of-concept with Apptronik and Mercedes-Benz, Apollo 2 humanoids using the model achieved a vendor-reported 96% success rate while sorting automotive parts on active production lines.

Why it matters
Robotics progress increasingly depends on generalization rather than teaching one machine one motion at a time. A model that transfers across objects, environments, and robot types could shorten the distance between a successful lab demonstration and useful work in production.

What’s next
Gemini Robotics 2.5 is rolling out to Google DeepMind’s robotics partners. Google is also releasing an adversarial robotics-safety benchmark and using Gemini 4 Argon to generate simulations, diagnose failures from video, and accelerate robot training.

⚙️ Arm Pushes More Intelligence Into Tiny Devices

What happened
Arm introduced the Cortex-M42 processor for low-cost, power-constrained intelligent devices. Arm says it delivers up to three times the vector-compute capability and twice the vector power efficiency of Cortex-M33 while supporting workloads such as sensor analysis, anomaly detection, condition monitoring, and embedded machine learning.

Why it matters
Physical AI cannot send every sensor reading to a large cloud model. Robots, industrial equipment, wearables, and connected devices need small processors that can recognize local conditions quickly while consuming little power.

What’s next
Cortex-M42 joins Arm Flexible Access and Total Access, programs now used by hundreds of silicon partners. Arm specifically points to local monitoring of robotic motors and joints as an example, while leaving safety-critical motion control to more specialized systems.

💡 Bottom Line

AI is becoming more specialized at the same time that it becomes more autonomous. Enterprises are wrapping agents in industry context and continuous controls, while researchers are measuring whether models can handle messy professional files and unpredictable physical environments. The durable advantage will come from combining intelligence with the context, boundaries, and infrastructure required to let it act reliably.

⚙️ Try It Yourself

See Which Model Handles Real Work Best

Today’s new Multimodal Knowledge Agent benchmark lets you inspect how frontier agents perform on professional tasks involving drawings, spreadsheets, scientific files, and other visual inputs.

Try this: Open the benchmark and choose one Example Task in the benchmark. Compare the submitted output from Claude Fable 5.2, Gemini 4 Carbon, and GPT-6 Astra. Then switch to Cost or Speed and compare the same models again.

Does the model with the strongest-looking output also have the best cost or execution time?

Insight: The best agent is increasingly workload-specific. Quality matters, but so do the cost and time required to finish the job.