Active

ARCHIE: exploring how reading and learning change when an AI reads along

Project: Archie: An AI-native Learning Companion that Strengthens Learner Agency, an intern project in LEXI, the internship at ASU's Learning Engineering Institute (LEI)
Role: Lead for design, UX and experimental code
started the repo, set its rules and maintain it
mentor
co-wrote the project guide
Team: Three leads, each heading a line: Nishad P (design, UX and experimental code), a PhD lead researcher (learning and pedagogy), a software lead (software engineering and operations)
3 interns
Timeline: August 2026 onward
Status: Experimental, in a private repo
the reading surface runs on placeholder data
Built with: React
TypeScript
JSON Schema
Markdown
a telemetry database
Claude Code

ARCHIE asks how reading and learning change when an AI companion reads along with a student, and whether its support strengthens the learner's judgment or substitutes for it.

Three of us each lead a line: learning and pedagogy, software engineering and operations, and mine, design, UX and experimental code. I started the repo, set up its rules and maintain it with our software lead; the interns build in it.

One interaction, every condition

Instead of everything at once, we took one part of learning, reading, and one interaction in it, the highlight, and we explore it under every condition we can think of. It's active exploration and experimentation.

01What comes in
A highlight, and what can come with it: a note, a reaction, a connection to another passage.
02The conditions
Who made it, the reader or the AI; what it says; how often highlights come over time; how closely they cluster in the text.
03What goes out, still being explored
When ARCHIE speaks up, how it says it, and on which surface, inline in the text or in a side panel. Each candidate cue is to be written twice, once leaving the thinking to the learner and once handing over the answer, so we can compare them.

The plan so far

Working plans, open to change:

A gate on every cue

Would this support the learner's thinking, or replace it?

The gate can only reject. Whether it's a model or a set of rules is still open.

The learner owns itWhat they need to see and verify is kept as plain Markdown files they own; the rest goes into a database. Agents only read; one writer writes. When it ships, the harness goes to the learner's machine too, and only the models run remotely.
Deciding when to speakFor now, a plain function reads the learner's policy file, which they can read and reset. A model can take that job if it needs to.

Reflection

ARCHIE is a cousin of my research effort LEAD and my filmmaking system LIVE: all three stem from harness engineering, the rules and tools agents work through, and collaborative work between people and agents.