Agents Get Identities. Governance Gets Real. Infrastructure Expands.

What happened
Microsoft introduced Scout, an always-on “Autopilot” agent for work that operates across Teams, Outlook, OneDrive, and SharePoint, with early access gated through its Frontier program; Microsoft is starting with a desktop preview before pushing toward a broader cloud version.

Why it matters
This is Microsoft’s clearest move yet from chat-based assistance toward a persistent enterprise agent with its own identity, policy controls, and visibility layer. In other words, the company is trying to turn “AI helper” into a real digital worker, but one that IT can still trace and govern.

What’s next
Microsoft says Frontier organizations can test Scout now, and its stated roadmap points toward a broader cloud-based rollout after these early deployments. The bigger test is whether enterprises trust it with higher-stakes background work without creating new security headaches.

What happened
Microsoft introduced Agent Control Specification in preview, letting teams define what agents may do, what requires human approval, and what evidence must be logged. Microsoft says ACS is meant to work across Foundry, Microsoft Agent Framework, and LangChain, while Microsoft is also shipping support across additional SDKs and tool stacks.

Why it matters
One of the biggest blockers for enterprise agents is not capability but control. Microsoft is trying to turn agent governance from scattered prompt rules and custom code into a reusable control layer that security and compliance teams can actually audit.

What’s next
ACS is launching in preview alongside broader tracing and evaluation tooling in Foundry. If developers adopt it across mixed agent stacks, this could become one of the first serious attempts at a cross-framework policy standard for production agents.

What happened
Microsoft launched seven new in-house MAI models at Build, led by MAI-Thinking-1, a 35-billion-active-parameter reasoning model, plus new image, voice, transcription, and coding models. Microsoft says MAI-Thinking-1 was trained from scratch on clean data without distillation from third-party frontier models, and several of the new models are already available in Foundry and MAI Playground.

Why it matters
This is a strategic shift, not just a model release. Microsoft is signaling that it wants more independence in core AI supply, more leverage in enterprise pricing, and more control over how models get adapted to customer workflows.

What’s next
Microsoft says MAI-Thinking-1 is open to select early partners, while Frontier Tuning is also opening to select customers so enterprises can train domain-specific variants inside their own environments. That means the next battleground is not only model quality, but who owns the tuned model, the workflow data, and the economics.

What happened
OpenAI launched “Codex for every role, tool, and workflow,” adding six role-specific plugins, Sites for shareable interactive work products, and annotations for in-place refinement. OpenAI says Codex now has more than 5 million weekly users, with non-developers making up about 20% of usage and growing more than three times as fast as developers.

Why it matters
This is Codex moving beyond engineering into mainstream enterprise workflows. OpenAI is no longer selling just an AI coding assistant; it is trying to make Codex the agentic layer that sits on top of documents, dashboards, data tools, and work artifacts across the company.

What’s next
OpenAI says more role-specific plugins are coming, including corporate finance, private equity investing, marketing strategy, strategy consulting, and legal. If that rollout lands, Codex starts to look less like a developer product and more like a horizontal operating layer for white-collar work.

What happened
Anthropic is expanding Project Glasswing and access to Claude Mythos to 150 organizations across more than 15 countries, extending its reach into critical infrastructure sectors such as power, water, healthcare, and communications. The broader rollout includes countries such as Canada, Australia, New Zealand, France, Germany, Japan, and South Korea, with named new participants including Samsung, Okta, SK Hynix, Intercontinental Exchange, and NATO.

Why it matters
This is one of the clearest examples yet of frontier AI being deployed first as defensive infrastructure rather than consumer software. Anthropic is effectively arguing that highly capable cyber models should be staged through controlled enterprise and national-security channels before anything resembling broad release.

What’s next
The immediate phase is more partner usage and more vulnerability discovery across a wider geographic footprint. The larger question is whether this becomes the template for other high-risk frontier models: controlled deployment to defended sectors first, broader availability later if safeguards hold.

What happened
Microsoft announced Majorana 2, a quantum computing chip whose materials were optimized by the company’s own AI tools. By switching from aluminum to lead—an unusual choice in chip fabrication—the team claims a 1,000‑fold improvement in certain performance metrics. Microsoft says commercially useful quantum machines could arrive in 2029, matching IBM’s timeline.

Why it matters
Using AI to design hardware highlights a new virtuous circle: AI advances enable better chips, which in turn accelerate AI. If validated, Majorana 2 could tilt the quantum race by making stable qubits sooner. Skeptics note that Microsoft has yet to release full data and has faced past criticism over reproducibility.

What’s next
Physicists and regulators will demand independent verification. Meanwhile, Microsoft must translate lab claims into scalable systems while fending off IBM, Google and Chinese rivals.

What happened
Microsoft unveiled Project Solara, a chip-to-cloud platform for “agent-first” devices, along with desk and badge reference concepts. Microsoft says Solara is built for enterprise manageability, multiple agents, biometric access, and “just-in-time” interfaces that adapt across form factors, while Pilot activity is planned with companies including AccuWeather, Best Buy, CVS Healthcare, and Target.

Why it matters
The important move here is not the demo hardware itself. Microsoft is trying to define the operating system, security model, and management layer for a future class of always-available AI devices that live outside laptops and phones but still connect back to enterprise identity and policy.

What’s next
Microsoft says the current badge and desk devices are reference designs rather than products, so the near-term test is whether partners actually build on the platform. If they do, Solara could become an early template for how physical AI gets deployed at work: managed, narrow, ambient, and agent-native.

What happened
At Computex 2026, Intel executives laid out a roadmap for “physical AI” devices and infrastructure. The company boasts over 4,000 edge partners and 100,000 deployments and plans to extend its 18A silicon into robots, autonomous machines and other AI‑powered form factors. Intel also introduced hybrid computing solutions—allowing sensitive inference on devices and broader tasks in the cloud—and launched Xeon 6+ processors with 288 energy‑efficient cores for agentic workloads. Executives argued agentic AI requires 1:1 CPU‑to‑GPU ratios and rack‑scale systems.

Why it matters
By embracing physical AI, Intel is repositioning itself beyond PCs and servers. Its emphasis on CPU‑heavy architectures for agents could reshape data‑center design and give it an edge against GPU‑dominated rivals. Partnerships with Foxconn, SambaNova and others indicate broad industry backing.

What’s next
Expect new AI‑enabled PCs, edge devices and robotics platforms built on Intel’s chips. The company’s rack‑scale “Vector Core Compute” offering will test whether enterprises are ready to invest in specialized agentic infrastructure.

The industry is rapidly building the operating system for autonomous work. Persistent agents, portable governance standards, defensive AI swarms, domain-specific copilots, agent-native devices, and physical AI infrastructure are all converging toward the same outcome: software that can act independently while remaining observable, controllable, and accountable. The next competitive advantage will come not from deploying a single powerful model, but from orchestrating entire ecosystems of agents, policies, devices, and infrastructure at scale.

Build a simple digital coworker using OpenAI Codex or Microsoft Copilot Studio to handle a real task—meeting notes, prospect research, report generation, or inbox triage. Then define three rules: what it can do automatically, what requires approval, and what must be logged.

Finally, stress-test the workflow with unexpected requests and permission boundaries.

Today's stories suggest the next AI advantage won't come from building smarter agents—it will come from governing, securing, and supervising them at scale.