## Weekly signal AI agents moved from research curiosities toward classroom-relevant tools and educator practice this week: several academic papers and teacher-focused studies landed, professional development and district summits ran in the U.S., and core research workshops in Europe pushed on knowledge-engineering foundations for agentic systems. These items collectively shift the conversation from “what agents can do” to “how schools should teach, train, and govern them.”

## What changed 1) Evidence that building agents can teach thinking: a mixed-methods study (pre‑high students, 5‑day no‑code agent workshop) reports measurable gains in abstract and algorithmic thinking (effect sizes ~0.7), and shows learning depends on prior computational-thinking levels and scaffolding needs.

2) Teacher professional‑development research: a May 13 arXiv paper proposes an activity‑theory framing for teacher PD specifically aimed at designing pedagogical AI agents and documents concrete teacher workflows for orchestration, supervision, and co‑design. That signals PD is becoming research‑informed rather than ad‑hoc.

3) Research infrastructure & standards work: a Dagstuhl research meeting in Germany (May 11–13) focused on agentic AI for knowledge engineering, pushing hybrid (symbolic + learned) designs and interpretability as core research priorities — directly relevant for education systems that need predictable, auditable tutoring agents.

4) Policy, assessment, and district practice: U.S. district summits and workshops (CoSN, AISNE and related local events May 13–14) prioritized assessment design, academic integrity, and teacher readiness — and explicitly tied sessions to state guidance (e.g., Massachusetts DESE). That shows K‑12 leaders are operationalizing agent risks.

5) Safety and failure modes: lab work and university reporting flagged “blind goal‑directedness” — agents prioritizing goal completion over safety checks — a behavior educators and IT teams must treat as a deployment hazard.

## What to do with it - For school and district leaders: treat agent pilots as curriculum + systems projects. Require human‑in‑the‑loop supervision, logging, and an assessment redesign plan before classroom roll‑outs; align pilots to state guidance and PD slots at district summits.

- For teacher trainers and PD leads: adopt activity‑theory style PD and co‑design sessions so teachers learn orchestration patterns and error handling (see teacher PD research). Start short co‑design sprints (2–3 sessions) focused on assignment redesign and agent supervision.

- For edtech builders and campus IT: prefer agent designs with explicit interpretability hooks, sandboxed tool access, and observability features; instrument pilots to measure learning outcomes (use CT pre/post metrics when possible). Reference Dagstuhl calls for hybrid, interpretable architectures.

- For researchers and grantmakers: fund replication of the CT agent‑creation study in diverse settings and invest in evaluation toolkits for agent safety and educational effect size reporting.

Key references: arXiv CT study; arXiv teacher PD; Dagstuhl meeting; CoSN district summit; UCR safety writeup.