Rethinking EdTech: How Multi-Agent AI Can Support ADHD Learners

Attention-deficit/hyperactivity disorder (ADHD) is characterised by differences in the brain’s prefrontal cortex, which affects executive functions like working memory, planning, impulse control, and task-switching. In a school environment, this creates a frustrating double burden: the students who most need structured revision are therefore the least cognitively equipped to structure it independently.

While Large Language Models (LLMs) and intelligent tutors have transformed education, a single monolithic AI tutor cannot simultaneously provide deep subject-matter expertise, monitor cognitive load, and offer metacognitive scaffolding for an ADHD learner. Furthermore, single-LLM sessions are inherently stateless, meaning they cannot maintain the persistent memory needed to implement spaced retrieval schedules or track longitudinal error patterns.

The “Centaur” Solution: A Team of Specialised Agents

In a 2005 freestyle chess tournament, the winners were not grandmasters or supercomputers, but two amateur players using three ordinary computers. As Garry Kasparov concluded, a weak human player plus a machine and a better process is superior to a strong player plus a machine and an inferior process. The architecture of collaboration matters more than the baseline strength of either component.

Applying this “centaur” model to education, this research proposes shifting the cognitive load from the learner to an ensemble of specialised AI agents. The architecture includes four distinct roles

• The Subject Tutor: Delivers curriculum-aligned content mapped to past paper questions.
• The Examiner: Administers retrieval practice using real mark schemes to counteract self-monitoring deficits.
• The Metacognition Coach: Provides scaffolding for task initiation and self-regulation.
• The Progress Tracker: Maintains a persistent learner model to offload revision management, such as planning and spaced retrieval scheduling.

Inside the Pilot Study and Measuring Success

To see how this architecture performs, a pilot study was conducted with “Student A,” a Year 11 learner with suspected ADHD. Over five days, the student completed seven tutoring sessions totalling 4.8 hours across Mathematics, Chemistry, History, Design Technology, and English.

For an AI tutoring system to be truly effective, success requires measurable academic and cognitive improvement. The primary objective is to target a reduction in the “grade gap” from an estimated baseline of two grades between supported and unsupported work, down to one grade or less.

The system’s success is tracked through several key quantitative measures:

• Exam Score Differentials: Comparing the student’s performance on past exam questions when supported by the multi-agent system versus their performance during unaided study.
• Self-Correction Frequency: Monitoring how often the student catches their own mistakes is an indicator of strengthening metacognition.
• Error Classification: Using NLP to detect and classify error patterns such as procedural, conceptual, or careless.

Decoding the Data: Text Analytics in Action

During the pilot study, the interactions generated 182 timestamped messages—51 from the student and 131 from the assistant. To evaluate these transcripts objectively, the research employed specialised text analytics methods.

• Tracking Emotions with NLTK and VADER: The Natural Language Toolkit (NLTK) was used to perform computational analysis on the corpus, including VADER sentiment analysis. Interestingly, VADER classified 70.6% of the students’ messages as neutral. This happened because educational dialogue—where answering a factual question correctly is highly engaged but linguistically neutral—falls outside VADER’s training distribution. Relying on keyword-based detection instead proved that frustration markers appeared in only 2.0% of messages, despite the student making 28 errors.
• Measuring Cognitive Load with Textstat: To prevent the student from becoming overwhelmed, the textstat library was used to compute Flesch-Kincaid grade levels. The AI pitched its language below formal GCSE level, achieving a mean Flesch-Kincaid grade of 6.1, deliberately reducing cognitive load.
• The Vocabulary Paradox via Tokenisation: Tokenisation revealed that the student had notably high vocabulary diversity, with a type-token ratio of 0.727. This proved their difficulty was not a lack of linguistic capability, but the executive function required to deploy it. Conversely, the AI’s vocabulary diversity was deliberately low (0.316). This repetition provided the predictability and routine necessary to compensate for ADHD executive dysfunction.
• Spotting Good Teaching: The transcripts were analysed using keyword pattern matching against curated indicator phrases. This demonstrated that the AI agents utilised formative assessment in 37.4% of their responses, confirming active pedagogical engagement. Furthermore, the student demonstrated a self-correction rate of 28.6%, indicating emerging metacognitive awareness without explicit prompting.

By moving away from monolithic chatbots and toward a coordinated, multi-agent ecosystem, educational technology can fundamentally shift to systems that adapt their internal structure to the learner’s neurological profile.