## Weekly signal

The education-and-learning signal for May 4–12, 2026 is that agentic AI is becoming operational infrastructure, not just a classroom discussion topic. The most useful developments were about workflow adoption in education workspaces, new hands-on curricula for building agents, faculty reassessment of assessment design, and security guidance that schools can actually use. Coverage is based on live search through May 11, 2026; no substantive May 12-specific items were visible yet.

The week also showed a split in the market. On one side, platforms and training providers are making it easier for students, faculty, and staff to build agents that use tools, manipulate data, and produce interactive outputs. On the other side, cyber agencies are telling organizations to slow down when agents touch sensitive data or critical systems. For education leaders, that is the right tension: agentic AI is useful enough to pilot now, but only if permissions, logs, evals, and human review are designed in from the start.

## What changed

OpenAI made spreadsheet-native AI more relevant to education operations. ChatGPT for Excel and Google Sheets became available globally for Enterprise, Edu, and K-12 workspaces. The release brings ChatGPT into a spreadsheet sidebar for building, updating, explaining, and reviewing multi-tab files, with support for Skills and apps where available. OpenAI also notes workspace controls such as RBAC, data and inference residency where available, Enterprise Key Management, and Compliance API coverage. The same release-notes page added global admin-console areas for Analytics and Agents, including views into active users, message activity, GPTs, projects, skills, tool interactions, connector interactions, workspace health, agent activity, connected apps, memory files, schedules, and agent run analytics.

For schools and universities, the near-term value is not “autonomous teaching.” It is operational leverage: enrollment trackers, financial-aid checklists, departmental budgets, research administration, curriculum mapping, assessment spreadsheets, student-success triage lists, and institutional reporting. The risk is that spreadsheet agents can quietly become decision-support systems. If a model recommends who needs outreach, which grant line item is wrong, or which student is off-track, the workflow needs provenance, review, and correction loops.

DeepLearning.AI released a new builder-focused course, “Build Interactive Agents with Generative UI.” The course teaches developers to build full-stack agent apps that go beyond plain text by connecting LangChain and Google ADK agents to React frontends with CopilotKit and the AG-UI protocol. It covers controlled, declarative, and open-ended generative UI patterns, including agents that render charts, cards, forms, and canvas-style shared workspaces.

This matters for learning products because many education tasks are visual and iterative. A tutoring or coaching agent should be able to expose a concept map, rubric, graph, simulation, worked example, or editable plan. A faculty-support agent should let instructors inspect generated assessments, align them to outcomes, and revise them in place. The practical shift is from “chat as the interface” to “agent plus shared workspace.” Builders working in edtech should treat generative UI as a core design pattern, especially where learners need to see intermediate reasoning artifacts without being overwhelmed by raw chain-of-thought.

Agentic AI is also becoming formal curriculum. In India, IIIT Hyderabad’s Division of Flexible Learning opened applications for a 12-week online certificate, “Engineering Agentic AI Systems: From Concepts to Practice.” The program is explicitly engineering-first: foundations of LLMs, RAG, autonomy, reasoning and planning; architectural patterns and multi-agent orchestration; implementation with tool use, memory, MCP, A2A, and agent frameworks; and evaluation, deployment, monitoring, AgentOps, and maintainability. The course begins June 29, with applications open from April 29 through June 15.

That program is a good example of where serious agent education is heading. The differentiator is not another prompt-engineering module. It is lifecycle coverage: scoping, architecture, implementation, testing, deployment, monitoring, and operations. Schools building similar courses should include failure modes, sandboxing, observability, evaluation datasets, cost controls, permission design, and human-in-the-loop checkpoints.

In the U.S., Arizona State University’s “The Agentic Self” course completed its first semester. According to AfroTech’s report, will.i.am and FYI.AI taught 75 students across 16 class meetings, including undergraduate, graduate, and non-degree learners. Students built personal AI-agent projects focused on small-business support, veterans’ benefits, and African-language learning support, with the top projects presented to judges. The report says the course was created without computer-science or coding prerequisites.

The ASU example is important because it shows a different curriculum pattern: agentic AI as civic, entrepreneurial, and personal capability rather than only a computer-science specialization. For non-CS programs, the useful design is not “make everyone an AI engineer.” It is: identify a community problem, design a safe agent workflow, define data sources and boundaries, test it with users, and explain risks. This can fit business, education, public policy, health communication, language learning, and workforce-development programs.

Faculty practice is changing too. Oxford Brookes University in the UK hosted a May 6 session on “AI-Assisted Coding, Assessment Design, and the Agentic AI Era.” The session focused on how agentic systems capable of autonomous reasoning and task execution are reshaping teaching, coding practice, and assessment design. Participants discussed how assessments should evolve, what academic integrity means when AI-assisted development tools are normal, and how to design authentic, future-ready assessment strategies. The reported direction was away from traditional assessment models and toward critical thinking, problem framing, creativity, reflection, and applied problem-solving.

That is a useful signal for every institution. If an agent can plan, code, debug, retrieve, summarize, and format outputs, then assessments based only on final artifacts will become weaker. Stronger designs will evaluate the student’s framing, constraints, critique, iteration history, defense of tradeoffs, and ability to use agents responsibly. Assessment policies should stop treating AI use as a binary yes/no issue and start specifying allowed tools, disclosure expectations, process evidence, and unacceptable delegation.

The governance backdrop also sharpened. Australia’s Cyber.gov.au published the Five Eyes joint guidance “Careful adoption of agentic AI services,” co-authored by ASD’s ACSC, CISA, NSA, Canada’s Cyber Centre, New Zealand’s NCSC, and the UK NCSC. The guidance says agentic AI can automate repetitive, well-defined, low-risk tasks, but organizations should assess what could go wrong, maintain visibility, avoid broad or unrestricted access, and use agents only for low-risk and non-sensitive tasks where possible. It highlights privilege, design and configuration, behavior, structural, and accountability risks, and recommends progressive deployment, least privilege, secure defaults, sandboxing, human oversight, and continuous evaluation.

For education, this is not abstract cybersecurity advice. Student records, disability accommodations, safeguarding notes, financial-aid data, research data, and HR data are sensitive. An advising agent with too much access can create real harm even if it is “only” drafting recommendations. A classroom agent with email or LMS write access can mis-send messages, expose records, or change grades if permissions are sloppy. The guidance should become part of procurement, pilot approval, and data-governance reviews for any education agent touching institutional systems.

## What to do with it

First, build an agent inventory before scaling. Track every agent-like workflow in use: owner, purpose, user group, connected tools, data classes, permissions, model/provider, trigger conditions, logging, evaluation method, and rollback plan. Treat spreadsheet assistants, Slack agents, LMS automations, and no-code workflows as part of the same inventory.

Second, classify use cases by consequence. Low-risk pilots include meeting prep, public-course content transformation, non-sensitive spreadsheet cleanup, draft rubrics, practice quizzes, and staff knowledge-base search. Higher-risk workflows include advising, accommodations, grading, financial aid, student discipline, procurement, payroll, and research compliance. Do not let convenience move a pilot from the first group into the second without a new review.

Third, update curriculum around systems, not prompts. Agent education should cover tool calling, memory, RAG, UI, multi-agent orchestration, evals, observability, deployment, security, and human review. IIIT Hyderabad’s structure is a useful blueprint for technical learners. ASU’s project-based model is a useful blueprint for non-CS learners.

Fourth, redesign assessment now. Require students to submit problem framing, assumptions, prompts or agent-use logs where appropriate, critique of outputs, iteration notes, and final reflection. Grade the judgment around the agent, not just the artifact.

Finally, when building learning agents, invest in interactive UI. Text-only chat is too limited for many learning tasks. Use charts, forms, editable plans, concept maps, simulations, and side-by-side workspaces so learners and teachers can inspect and steer the agent’s work.