Platforms / Programming education

MIRA-Agent

A programming tutor that asks students to show what they understand before AI-assisted code can run.

Domain: Data science programming in Python
Learners: Undergraduate and graduate students
Gates: Concept check, dry run, and code check
Theoretical basis: Self-regulated learning: forethought, performance, and self-reflection

Why it exists

Working code is no longer reliable evidence of learning. A student whose code runs may never have understood what it does. MIRA-Agent studies the process that produces the code rather than the code alone.

MIRA, the AI mentor, does not write the solution. Before code can run, the student answers a concept question, predicts the output of key steps, and passes a check against the task. When a student pastes a block of code, MIRA asks about it before the student can continue. Students see their own working pattern on a five-dimension profile while they work.

What students see, and what the platform records

MIRA-Agent workspace with MIRA on the left and a Python editor on the right — Workspace
Students work through a multi-step assignment with MIRA beside a Python editor. Code runs only after the student answers MIRA’s concept question.
Records every message to MIRA, each gate attempt, and each code run.

Dry run prompt asking the student to predict the output of the code — Dry run
Before key steps run, students predict the output in their own words.
Records each prediction and the number of attempts.

Paste detection

When a student pastes a large block of code, the platform flags it and pauses authorship credit. MIRA then asks the student to explain the pasted code.

Records the size of each paste and how much the student changed it afterward.

Reflection

After the assignment, students see a five-dimension profile of how they worked, with advice for the next assignment.

Records the five dimension scores across the session.

What it measures

Five dimensions, grouped by the phases of self-regulated learning and updated after each interaction. The profile draws on seven behavioral sensors, including response timing, pauses, backspace rate, and paste events.

Planning

Forethought

Working out the concept before writing code, shown by passing MIRA’s concept questions.

Comprehension

Performance

Predicting what code will do and producing code that meets the task.

Persistence

Performance

Retrying with a real change after a failed check or an error, rather than pasting a fix.

AI Collaboration

Self-reflection

Using MIRA to reason through a problem instead of asking it for answers. Drops when hints are requested without effort.

Authorship