Agentic AI in Education & EdTech

1 · From answer-bot to learning-loop agent

The learning lifecycle -- Assess fi Plan fi Teach fi Practise fi Grade fi Intervene fi Report -- has spent a decade being 'personalised' by adaptive-content engines that still left planning, marking and pastoral judgement on the teacher. Agentic AI changes the unit of automation: instead of serving the next worksheet, an agent now runs the whole loop -- diagnosing a misconception from a wrong answer, generating a tailored explanation, setting practice at the right level, marking it, and surfacing a struggling student to the teacher before the test does. The difference from earlier adaptive learning is the same difference seen across every vertical this series: the agent crosses steps and produces an outcome -- a lesson taught and a record updated -- not just a single

2 · What works today vs what stays teacher-owned

The reliable 2026 pattern: agents do the high-volume, low-stakes scaffolding; humans own anything that grades, flags or follows a child. The seam is the assessment decision -- exactly the line FERPA and the EU AI Act both The assessment seam: agents tutor, draft and flag -- but must NOT decide the grade, the placement, or the safeguarding outcome. Design the hand-off precisely at the line where a record

3 · Governance -- the minor-protection bind

Education carries a double regulatory load no other vertical shares in this form. In the US, FERPA protects the student education record and COPPA restricts data collection on under-13s -- so any agent reading or writing student data must scope access to legitimate educational interest and keep a disclosable audit trail. In the EU, the AI Act places systems used to determine access to education, evaluate learning outcomes, or monitor exam behaviour into Annex III high-risk, enforced from Aug 2 2026 (~T-55 days): risk management, automatic event logging over the system lifetime, human oversight, technical documentation and transparency to the learner, with €35M / 7% GTR penalties. The emerging standard mirrors the 'reviewer-of-record' pattern from legal (Day 73): a per-student reasoning log for every grade and every flag, plus a named teacher signing off on anything that scores, places or refers a child -- one record that satisfies EU oversight AND US fair-treatment duties at once. KYA (Day 54) = SPIFFE/SVID scoped read-and-propose (tutor / draft / flag, never grade:commit / place /

4 · Platform landscape

5 · Reference architecture for a learning agent

Combining the memory (Day 44), reliability (Day 49), identity (Day 54) and streaming-data (Day 55) patterns,

The one rule: the agent teaches, the teacher decides -- on every output that scores, places or refers a child. One trace satisfies EU AI Act Art. 12 logging, FERPA access audit and the duty of care

6 · Does it actually help? The 2 sigma question

The honest 2026 read: tutoring agents show real gains on well-scoped skills (languages, maths fluency, coding) and clear time savings for teachers on planning and first-pass grading -- but the full Bloom '2 sigma' effect remains unproven at population scale, and the risk that students outsource thinking rather than build it is the dominant pedagogical worry. Hence the convergence on Socratic-by-default designs (Khanmigo, Study Mode, Guided Learning all refuse to simply hand over the answer) and on teacher-in-the-loop as a product constraint, not a setting. The measurable wins live on the teacher side first: reclaiming hours from lesson prep and marking, which is also the lowest-risk place to start.

7 · Breaking this week

Anthropic confidentially filed for a US IPO (June 1 2026) at a reported ~$965B valuation -- on track to be the first AI company toward a trillion-dollar listing, ahead of OpenAI which is preparing its own confidential filing for a possible September debut. It rides Claude Opus 4.8 (shipped May 28, topping GPT-5.5 and Gemini 3.1 Pro on key benchmarks, with the new Dynamic Workflows tool spawning parallel subagents) at unchanged $5/$25 per M-token pricing. OpenAI shipped real-time audio + translation models for agents (live voice tutoring becomes practical) and is expected to release GPT-5.6 in June. Google pushed Gemini deeper into managed task flows; Microsoft unveiled its own MAI coding model at Build to cut reliance on OpenAI. The frontier is converging on agentic coding + orchestration -- and the public-market race is now on.

8 · Viral app of the day

GitHub Copilot App (agent-native desktop) -- GitHub shipped a standalone desktop 'control center' for agentic development this week, moving Copilot from in-editor assistant to a surface where you dispatch, watch and steer multiple coding agents at once. It lands amid a fortnight where a new agent repo went viral almost weekly -- OpenClaw (past 210K stars, fastest-growing OSS in GitHub history), Hermes Agent (100K+ stars, self-improving 'grows-with-you' memory + a new desktop app), and mattpocock/skills ('skills for real engineers', ~91K stars in a week). The education lesson is the same one that runs through this whole series: the most viral agents are the most autonomous, and many ship with no kill switch, no audit trail and no scoped identity -- the exact opposite of what an agent touching a child's record must have. The govtech and edtech job is to give students and teachers a sanctioned, governed agent before an ungoverned one fills the gap.