Agents Take Control. Identities Get Enforced. Trust Gets Tested.

What happened
Anthropic launched a research preview that lets its Claude chatbot operate a Mac like a human, clicking buttons, typing in fields and opening apps. The feature lives inside Claude Cowork and Claude Code and uses a layered approach—first trying direct connectors (Gmail, Drive, Slack), then a Chrome extension, and only last resort interacting with the user’s screen. Early testers report the tool works about half the time and struggles with complex multi-step tasks, highlighting that it’s still a research preview.

Why it matters
This is one of the first consumer AI agents that actually performs work rather than chatting. It thrusts Anthropic into a head-to-head race with OpenAI, Google and Nvidia to build agents that navigate existing software and underscores how quickly agentic AI is moving from demos to real workflows. It also raises privacy and compliance concerns because the agent can see anything on a user’s screen and the safeguards are imperfect.

What’s next
Anthropic plans to expand beyond macOS, integrate more deeply with its Dispatch mobile feature and improve reliability. As enterprises adopt such agents, regulators and customers will demand audit trails, identity management and kill switches to mitigate security risks.

What happened
Identity-security company Okta introduced a “secure agentic enterprise” framework and announced a new platform, Okta for AI Agents, set for release in April. The company notes that 88 % of organizations have experienced suspected or confirmed AI-agent security incidents, yet only 22 % treat agents as first-class identities. Okta’s blueprint asks three questions—Where are my agents? What can they connect to? What can they do?—and proposes registering agents in a central directory, enforcing least-privilege access via an Agent Gateway and providing a universal logout “kill switch”.

Why it matters
Existing IAM systems were built for predictable human users. Autonomous agents can spawn other agents and execute workflows continuously, creating a rapidly expanding attack surface. Treating agents as non-human identities brings visibility and continuous authorization into the core of security operations, closing gaps that attackers could exploit.

What’s next
The platform will enter general availability next month. Expect other IAM providers and managed-security firms to offer agent registries and real-time revocation services as enterprises adopt agents across business functions.

What happened
The Cloud Security Alliance revealed at RSAC that 73 % of organizations expect AI agents to be vital within a year but 68 % cannot clearly distinguish AI-agent activity from human actions. Respondents reported widespread use of agents for task automation, research and monitoring, often under borrowed identities; ownership of agent identity was fragmented across teams and governance mechanisms were being used as a stopgap for missing access controls.

Why it matters
As agentic AI proliferates, the inability to tell who—or what—is performing actions poses a governance and liability risk. The findings underscore the urgency for identity management and audit trails designed specifically for autonomous agents.

What’s next
Organizations are likely to adopt agent registries, continuous monitoring and role-based controls, while regulators may require auditable records of agentic activity. Vendors will need to embed identity and safety features into agent-building platforms to gain enterprise trust.

What happened
At Oracle AI World London, the database giant unveiled 22 pre-built Fusion Agentic Applications and an expanded AI Agent Studio. These agents are embedded across Oracle’s HR, finance, supply-chain and sales suites to handle tasks such as updating customer accounts, reconciling invoice disputes and flagging supply-chain bottlenecks. Oracle also introduced a Private Agent Factory with no-code builders, vector search, unified memory and deep data security to ‘future-proof’ its database for the agentic era.

Why it matters
By bundling specialized agents directly into its cloud applications, Oracle aims to leapfrog chatbots and offer orchestrated workflows that span departments. The unified memory and vector capabilities support complex queries and reduce integration work, positioning Oracle as a serious player in enterprise agentic AI.

What’s next
Oracle plans to roll out the tools in limited availability this year, measuring ROI for early customers. Competitors like Salesforce, SAP and Workday may follow suit with their own agentic suites as enterprises test whether these agents actually reduce manual work.

What happened
MariaDB announced it completed its acquisition of in-memory data-grid provider GridGain. The combined platform merges transactional and analytical processing with in-memory speed and built-in vector search, creating a unified data layer for autonomous agents. Analysts cited by MariaDB predict that 40 % of enterprise applications will feature task-specific AI agents by 2026 and warn that poor data foundations could cost enterprises 15 % productivity.

Why it matters
Agentic AI needs high-velocity data pipelines that support both real-time transactions and analytics. By integrating GridGain, MariaDB aims to eliminate the need for developers to stitch together separate databases, promising faster response times and simplified architectures.

What’s next
MariaDB will position the new platform for AI-heavy workloads, with customers like Hatch already testing it. Expect other database vendors to pursue similar acquisitions or partnerships to create AI-ready stacks.

What happened
Researchers at Georgia Tech unveiled the first generative AI for polymer design—a chemical language model that suggests polymer structures based on target properties. Trained on 20 million monomer pairs, the model generated candidate polymers that were validated in physical experiments, producing materials with tailored mechanical and thermal characteristics.

Why it matters
Generative models are moving beyond text and images into materials science. Automating polymer design could accelerate the discovery of lightweight, strong or heat-resistant materials for aerospace, automotive and consumer products, bridging digital algorithms with physical manufacturing.

What’s next
The team plans to broaden the training data and integrate multi-objective optimization. Industrial partners may adopt similar models, signaling a future where AI co-designs new materials alongside human scientists.

What happened
Databricks announced LakeWatch, a new AI security product, alongside an open, agentic security framework designed to monitor, govern, and protect AI systems. The platform focuses on securing data, models, and agent workflows across the AI lifecycle, with capabilities like real-time monitoring, policy enforcement, and threat detection.

Why it matters
As enterprises move from copilots to autonomous agents, the attack surface expands fast. Databricks is positioning security as a core layer of the agent stack—embedding governance directly into how agents access data, make decisions, and take action, rather than treating it as an afterthought.

What’s next
Expect a wave of “agent-native security” platforms that combine observability, identity, and policy enforcement into a single control plane. As adoption grows, security will shift from perimeter defense to continuous oversight of autonomous systems.

What happened
Spotify is beta testing “Artist Profile Protection,” letting artists approve or decline releases before they appear on their profiles, targeting AI-generated music misattribution.

Why it matters
This is a direct response to the flood of AI-generated tracks, signaling a shift toward platform-level controls and verification in the music industry.

What’s next
If successful, expect broader rollout and similar controls across other streaming platforms, plus industry-wide discussions on AI content rights.

What happened
Doss secured $55M in Series B funding to expand its AI-native inventory management platform, integrating with both traditional and AI-first ERP systems.

Why it matters
The rise of modular, agent-friendly infrastructure is reshaping enterprise software, as brands move away from legacy ERPs toward composable AI stacks.

What’s next
Competition will intensify among vendors rebuilding ERP for agentic compatibility, setting new precedents for enterprise SaaS.

What happened
ANELLO Photonics and quantum-control firm Q-CTRL announced a partnership to develop resilient navigation solutions for GPS-denied environments. By combining ANELLO’s silicon-photonics inertial sensors with Q-CTRL’s quantum-magnetometer technology, the companies aim to provide accurate positioning when GPS signals are unavailable or jammed.

Why it matters
GPS vulnerability is a critical weakness for autonomous systems. Hybrid quantum-photonics navigation could enable drones, aircraft and vehicles to operate safely without satellite signals, opening new markets in defense, logistics and aviation.

What’s next
The partners will prototype and test navigation units with commercial and government customers. Success could spawn an industry of quantum-enabled navigation devices and reduce reliance on GPS.

What happened
FANUC America announced a $90 million investment to build an 840,000-square-foot facility in Michigan, adding 225 jobs and expanding its U.S. footprint to 3 million ft². The new plant will support growing demand for industrial robots across automotive, warehousing and electronics industries; FANUC has invested nearly $300 million in U.S. facilities since 2019.

Why it matters
Demand for automation continues to surge, and FANUC’s expansion underscores confidence in U.S. manufacturing for physical AI. A larger domestic facility strengthens supply chains and helps customers avoid long lead times or import disruptions.

What’s next
Construction begins this year. With supply-chain resilience a priority, rival robotics manufacturers may announce similar expansions to capture a share of the growing automation market.

What happened
Agile Robots announced a partnership with Google DeepMind to integrate Gemini Robotics models into its systems, targeting deployment in manufacturing, logistics, and more.

Why it matters
This marks a major step in merging advanced AI with physical robotics, accelerating embodied AI deployment in real-world industrial environments.

What’s next
Gemini-powered robots will roll out in industrial settings, with ongoing data collection to refine AI models and spur further industry partnerships.

As agents move from assistants to actors, identity becomes the control plane for everything they touch. The winners won’t just build smarter agents—they’ll build systems that know exactly who (or what) is acting, what they can access, and when to shut it down.

Want to see how “agent + identity + control” actually plays out?

→ Go to https://chat.openai.com or https://claude.ai/
→ Ask: “Act as an autonomous agent that can complete tasks using tools. What permissions do you need before starting?”
→ Then tighten the constraints: “Now only allow read-only access. What changes?”

You’ll notice something quickly:
the behavior of the agent is entirely shaped by what it’s allowed to do.

That’s the control plane.