News: OpenAI Presence, Claude Tag and Gemini Education move AI into governed institutional workflows

This product-news roundup tracks a new institutional question: not merely whether an AI can complete a task, but how an organization can decide what the system may access, which actions it may take, and when a person must intervene. OpenAI's July 22 launch of Presence packages policies, permissions, evaluations, guardrails, approved actions, and escalation rules around voice and chat agents. Anthropic's Claude Tag brings a shared, proactive Claude into selected team channels. Google's current Gemini for Education offer pairs teaching, learning, and work purposes with administrator-managed access and visibility. Together, the products make governance part of the interface rather than a policy document that sits elsewhere.

OpenAI describes Presence as an enterprise product for deploying agents into customer and internal workflows. Each deployment begins with a defined job and receives only the knowledge and system access needed for that job. The organization specifies what the agent can do, what requires approval, and when a person should take over. Before launch, simulations and graders test common requests, edge cases, tool use, policy compliance, and escalation. After launch, production sessions and handoffs provide evidence for improvement, while Codex can propose changes that teams test and approve.

That design is relevant to education even though Presence is not marketed as a classroom tutor. Universities and school systems increasingly consider agents for IT support, admissions, student services, library help, and administrative enquiries. In those settings, a fluent answer is not enough. Institutions need to know which record was consulted, whether the user was verified, what action occurred, and why a case was escalated. OpenAI's July 15 release notes add a directly education-specific signal: apps with sync now support ChatGPT Edu workspaces that use Enterprise Key Management. Connected data can therefore sit inside stronger workspace controls, but institutions still need to minimize access, test failure cases, and keep consequential decisions with authorized people.

Anthropic's Claude Tag takes a different route to institutional AI. In its current beta for Claude Enterprise and Team customers, a team can invite Claude into selected Slack channels, connect approved tools, data, and codebases, and delegate multi-step work in a shared thread. Anthropic presents this as an evolution of Claude Code: the agent can break work into stages, operate asynchronously, schedule tasks, retain channel-scoped context, and, when enabled, flag relevant developments proactively. Administrators can separate identities, memories, information, and tools by channel so that one team's Claude does not expose another team's context.

For educational organizations, that shared visibility may be valuable. A curriculum team could see how an agent assembled resources, an IT group could review a troubleshooting trail, or a research lab could coordinate routine checks. The risks are equally concrete. Ambient monitoring can collect more context than participants expect, long-lived memory can preserve mistakes, and a visible thread may create false confidence that someone has reviewed every action. A safe pilot should define the agent's role, restrict connectors, label automated contributions, set retention rules, and name the human owner for approvals and corrections.

Gemini for Education provides an explicitly educational comparison. Google's current product page describes Gemini as supporting three purposes: teaching, learning, and work. Educators can plan lessons and differentiate materials; learners can deepen understanding, practise for assessments, and receive writing feedback; staff can use a managed assistant within an education environment. The institutional layer includes administrator control over access, usage visibility, Vault search and export, and enterprise-grade data protection. Google also distinguishes the standalone Gemini for Education service from Google AI Pro for Education, which expands access to Gemini, NotebookLM, and AI inside Workspace apps.

For Hong Kong schools and universities, the combined product signal is that AI adoption is becoming an identity, permission, memory, and escalation problem as much as a model-selection problem. A responsible evaluation should begin with a specific educational or service purpose, map every data source and tool, test refusal and handoff behaviour, log consequential actions, and measure whether people can understand and challenge the agent's work. Vendor announcements demonstrate product direction, not independent evidence of educational benefit or regulatory compliance in a local setting. The opportunity is to make routine support more responsive while keeping teachers, students, and authorized staff in control. The danger is to mistake configurable guardrails for completed governance.