The Rise of the AI Agents
TL;DR: AI agents are here! And they’re here to stay.
TL;DR: AI agents are here! And they’re here to stay.
Today we’re going to spill the tea on agents - think of them like your robot butlers - and they’re quickly becoming phenomenal teacher aides and operational workhorses in education. We’ll talk about how they work, what they can and cannot do for you, and ideas for how to get the most out of them in your educational workflows. We’ll highlight three signals of progress, a field note from the frontlines of education, a playbook for action, and a hype-check on what’s real (spoiler alert: AI agents won’t be solving education’s biggest challenges overnight, but they can tackle specific pain points, at scale.)
How AI agents work
AI agents, essentially virtual machines that can perceive their environment, reason about what to do next and reliably act toward specific goals, have been around for a while, but until last week, they were only accessible to the crowned handful of engineers who knew how to build them from scratch.
With the launch of ChatGPT Agents, Google GEMs, IBM Agent, and Microsoft AutoGen, agents are now accessible to the masses, making it possible for any organization to build and automate flexible and adaptable workflows. Think Zapier Zaps with extra sparkle and without the complicated wiring.
“The key is making agents both highly capable and user-steerable.”
What this means for education
Teaching at the Right Level (TaRL)
What the evidence says: Until today, one of the biggest levers, and challenges, in education has been real-time adaptation and awareness, i.e. TaRL.
We have plenty of content and curriculum, we understand TaRL, and we know that teachers need support to assess and adapt their lessons to their students’ learning gaps of the moment.
Long before generative AI, simpler “teacher assistance” technologies were already yielding clear gains.
Ghana’s Teacher Community Assistant Initiative (TCAI) found that hiring local classroom assistants for short remedial sessions and helping teachers group pupils by current skill level led to clear, lasting improvements in children’s reading and math (Duflo, Kiessel, & Lucas, 2024).
In India, we saw Mindspark deliver significant gains for low performers using adaptive math software (Muralidharan, et al, 2019).
What the opportunity is: Today’s AI agents can turbo-charge this early work by dynamically adjusting to each learner’s and teacher’s needs in real time.
Imagine Amira’s reading coach surpassing a human tutor’s oral reading gains by employing personalised and real-time contextualised stories, classroom-awareness, and easy-to-read summaries for teachers, written in plain language.
Imagine an “assessment-to-groups” agent that auto-generates 10-item assessments, auto-scores paper captures using tools like LearnLens, then suggests 3 skill-groups and group-matched lesson plans for the week, all within 30 minutes at the top of each week.
Imagine a timetable agent that ingests existing timetables, reserves three short TaRL blocks, shares rosters and daily timetables with the teacher over WhatsApp, and tracks who moved up, who moved down, and who needs extra attention for the week.
Why this works: These tasks are deterministic and repeatable, ideal for orchestration. TaRL’s effect depends on execution fidelity more than fancy modelling. Agents handle the coordination so teachers can teach.
Structured pedagogy
What the evidence says: “Structured pedagogy”, i.e. packaged teacher guides, aligned student textbooks, combined with consistent coaching is one of the most cost-effective ways to raise FLN outcomes, especially in low-resource/low-performing contexts. But maintaining alignment across programs and shifting country strategies, alongside watching each learners’ pace, trying to scale to multiple countries, conducting assessments for multiple programs, and updating coaching scripts is a publishing, printing, and logistics nightmare.
What the opportunity is: AI agents can offer a balm to Ministries of Education or county departments of education trying to keep up, while holding costs down. Agents may be able to deliver everything (that SP requires), everywhere, all at once.
Imagine a curriculum agent that takes a national syllabus and generates weekly guides, worksheets, and assessments localized to varying teacher levels of “English-medium” and pace, sending short, actionable updates over WhatsApp directly to teachers, and regenerates as and when the ministry updates the syllabus.
Imagine a “print-friendly publisher” agent that materially cuts costs by maintaining high teacher-friendly readability while shrinking page count and ink coverage, then runs a cost model comparator to determine the best place to print.
Imagine a choose-your-own-adventure agent that allows a teacher to select their preferred delivery channel (WhatsApp, print, pdf, etc.), which ingests the generic material and publishes to each teacher according to their preferred method, reducing total program costs as teachers become more and more mobile-first.
Why this works: Faster, cheaper, and more automated localization, personalization and digitization is now possible without extra wiring, hiring and enormous, sometimes wasted, printing costs. With agents, various workflows can interact and make semi-autonomous decisions to give incredibly powerful outputs.
Field note
Victor Appiah, a school leader in Kumasi, rolled out Rising Academies’ agentic Rori math chatbot to his Grade 3 class. After a term, students’ average scores rose from 74% to 82%, and in a controlled pilot, Rori users beat the control group by 11 percentage points. Students loved the instant feedback and Victor loved the time saved and the equity boost for slower learners. Parents are even buying cheap smartphones so their kids can keep using the tutor at home - proof that when a tool is low-bandwidth, curriculum-aligned, and teacher-endorsed, community trust follows.
Playbook for action
For funders – Back evidence-first pilots where the incentives of the tool are aligned to making the AI work better/harder. Small grants can test an AI grading bot or WhatsApp tutor in 20–30 schools with a quasi-experimental review, but what’s even more exciting right now is finding solutions where there is real upside when the AI works. Favour open-source or easily localisable tools (Kiswahili, Hindi, Wolof) and budget for teacher training plus a policy pathway so ministries can adopt if it works.
For operators – Pick one pain point to deploy an agent to. If uptake on your coaching feedback is dragging, try an agent that turns raw scores into ready-to-send, bite-sized, easily digestible, somewhat teacher-personalized insights. If differentiation has been tough, schedule a short after-school AI tutor session for remedial groups on an off-the-shelf, lightly customized tutor agent, and measure time saved and user engagement. Recruit “AI champion” teachers, plan low-tech fallbacks, and document lessons learned for peers.
For researchers – Embed experiments now. Test whether an AI agent–generated and -managed survey assistant boosts productivity and analysis quality. See if deploying mixed-methods research (ie. class observations + log data) will reveal why a tool works (or doesn’t).
Hype-check / Big no-no’s
Thinking that all “AI” is created equal. A fancy agent might ingest a curriculum and compose a decent lesson plan, but it might also get facts wrong or suggest activities that don’t fit the local context (one review of OpenAI’s latest Study Mode tool praised the innovation but noted at least five critical gaps limiting its pedagogical usefulness).
Agents acting purely on their own, with zero human/adult review before use with students – hallucinations and misinformation are still very real, especially in low-resource languages and cultural contexts outside of model data corpus.
Deploying agentic-only or online-only solutions in poor connectivity environments, causing unreliability and frustration to end users.
High-compute AI agents that lack a sustained way to recover costs.
In conclusion: We suggest you treat agents as augmenters, not teacher replacements; remember that solid pedagogy, training, and governance beat shiny algorithms and agents every time.
… and in other news
There was a wave of AI-learning tools released by Big Tech in late July/August: Claude’s learning modes were made accessible to all users, Google released Gemini for Education and subsequently Gemini Guided Learning (and also announced a $1 billion funding pledge for AI and education initiatives), OpenAI launched ChatGPT Study Mode. While these tools are online-only, there may be opportunities for the Foundation to influence how AI reaches low‑resource classrooms, including how these can be adapted or integrated with offline-first tools to leverage local teaching and learning materials.
Google released Gemma 3 270M, a highly compact, power efficient, open-weight model ideal for on-device deployment and privacy-sensitive applications. Compact AI models are becoming increasingly important in the AI ecosystem—especially for education, development, and deployment in resource-constrained contexts because of their lower computational requirements, faster response times, and on-device, offline capability.
There is growing concern on AI personality (read Ethan Mollick on AI personalities, UNICEF on AI bots and children, and Derek Thompson on AI as a social technology). Solution developers and education stakeholders must consider risks to young users, who could face increased risks of misinformation and emotional harm from technologies that are perceived to be objective and reliable, and can pose serious privacy risks, e.g. collecting data from children without proper guardrails.
Contributors to this issue of the newsletter include Shabnam Aggarwal, Robin Horn, and Ayesha Khan.
