The company says the release adds separate identity controls for each AI worker and workplace, a standard that lets AI tools connect to other software governance, a sandboxed runtime, continuous evaluation, and 40+ built-in guardrails. WSO2 cites Gartner’s prediction that the average Fortune 500 enterprise will have more than 150,000 AI workers in use by 2028, while only 13% of organizations believe they have the right AI-worker governance today. WSO2 announcement The important move is separating control from the individual agent. If governance lives inside each model vendor or application, every model change creates another rebuild. WSO2 is betting that the durable enterprise layer is identity, policy, evaluation, and suspension across a mixed estate, not another software environment where AI workers run. The market is moving toward a common accountability surface because number of AI workers will outrun the ability of point tools to explain who acted, with which permission, and at what cost. The decision for operators is to define the shared management layer before too many disconnected AI workers becomes the operating model. WSO2 announcement

The company says it assembles the context needed for a specific job instead of presenting one static map, with new AEC capabilities for automatic context enrichment, project-workflow actions, task-based AI workers, and a connector to Autodesk Forma. Egnyte announcement The AI model is not the hard part when the system cannot tell which document, person, project, or permission matters. Egnyte is trying to own the business-context layer that turns a generic answer engine into a work system. That shifts value toward the organization that can connect verified business facts across repositories without flattening permissions or losing provenance. The strategic risk for enterprises is allowing every department to rebuild context separately, producing a dozen inconsistent versions of what the business knows. Egnyte announcement

Workiva also announced three AI systems that complete multi-step work for Bureau of Economic Analysis surveys, U.S. Census surveys, and Country-by-Country Reporting, plus an automated testing solution for Internal Audit and GRC that coordinates evidence, attribute, and testing AI workers. Workiva says it serves more than 6,700 organizations and more than 85% of the Fortune 1,000. Workiva announcement This is a move from AI as a separate chat surface to AI inside a governed record of work. The advantage is not that a model can draft a filing; it is that the workflow can preserve evidence, traceability, and review while the system performs more of the repetitive work. That makes the application owner, not the model vendor, the likely winner in high-accountability domains. Every industry with audits, filings, or control testing will face the same packaging question: is AI a helper beside the process, or a participant inside it? Workiva announcement

The company says its software factories can run in cloud, on-premises, or fully fully isolated from outside networks environments, and that Factory Router reduced AI usage cost by more than 60% while maintaining frontier performance. Factory names Nvidia, Blackstone, Royal Bank of Canada, Palo Alto Networks, Adobe, and T-Mobile as customers building software factories with the platform. Factory announcement The funding is a market signal, not proof that autonomous coding has solved software delivery. Factory is selling a managed production system around coding AI workers: model choice, deployment location, governance, and measurement of value created by AI work. That is where the economics move when code generation becomes abundant. The constraint becomes review capacity, architecture quality, test coverage, and the ability to connect engineering output to business outcomes. CTOs should read the valuation as permission to test the operating model, not as permission to remove the human system around it. Factory announcement

NVIDIA reported that Lambda ran 19 nodes within a power budget typically allocated to 16 full-power nodes, increasing cluster-wide AI work completed 24% from about 4 million to 5 million units of AI output per second and improving performance per watt 23%. NVIDIA also says Vera Rubin and DSX MaxLPS can deliver up to 40% more GPU capacity within the same site-power envelope in the right environments. NVIDIA announcement The scarce resource is no longer only compute. It is power that can be converted into reliable, useful work. By tying infrastructure to AI that can complete multi-step work AI work completed, NVIDIA is pulling the data center into the business-outcome conversation. That will change capital planning: a facility with better power management can create more capacity without adding a new grid connection, while a facility that measures only peak speed can overstate its economics. CEOs and CFOs should ask for useful-work-per-megawatt metrics before approving the next AI buildout. NVIDIA announcement

The Models > Tuning workflow lets users configure Python code or model-based written definitions of what good work means, test reward logic against sample prompts, track training and evaluation metrics in real time, and test tuned checkpoints in Agent Studio. Google Cloud release notes This is a shift from consuming a general model to teaching a model what good work looks like inside a business. The a written definition of what good work means becomes a management artifact: it encodes quality, risk, speed, and policy choices that used to live in human review. That makes model improvement inseparable from operating-model design. Enterprises should resist tuning for a narrow benchmark and instead define rewards around the outcomes they are willing to defend. Google Cloud release notes