Nishad Patne

Projects / LEAD

Active

LEADlearning engineering architecture and design: researching learning interfaces, on infrastructure built for people and their agents

1 / 4The list, filtered by whose tags you're looking through; one source is contested, because two people disagree on its focus and both votes stand

Project
LEAD (learning engineering architecture and design), a research effort at ASU's Learning Engineering Institute (LEI)
Role
  • Research direction
  • designed and built the infrastructure: its policies for people and agents, its coordination and its interface
Team
Nishad P, with one teammate
Timeline
August 2026 onward
Status
  • In progress. Research questions at draft 3
  • the infrastructure is in use, in a private repo
Built with
  • Markdown
  • YAML and JSONL files in git
  • Node.js
  • React
  • Vite
  • TypeScript
  • Claude Code

Building a digital interface is cheap now.

I believe the next few months will bring more generative UI, more 3D built on WebGPU, and learning tools that feel more like games. LEAD is my research into what that means for education and learning. I set its direction and built the infrastructure it runs on, which is agentic UX in its own right: its design is a harness, the policies and commands that people and their AI agents work through.

01

32 sources and 188 thoughts

Before LEAD, what we read and thought lived in heads, browser tabs and DMs, out of reach of teammates and of AI. Since August, a teammate and I have collected 32 sources and 188 thoughts on them in LEAD, and our research questions grew out of them. LEI's postdocs and researchers are interested in working the same way.

  • 32sources

    Since August

  • 188thoughts

    LEAD

02

The harness is the design

The first version had grown into a project about itself: seven concepts, thirteen commands, an admin to approve tags.

I rebuilt it around one test: every rule protects against a real risk, or it goes.

What survived:

The library in the browser.

1 / 4The list, filtered by whose tags you're looking through; one source is contested, because two people disagree on its focus and both votes stand

  • One rule, enforced in the tools.Read anything, write only your own. Agents act on reasonable requests like "reorganise the tags for all", and a line in an instructions file can't stop them, so the commands can't reach anyone else's folder. An unclear request narrows to the agent's own person, and the agent says so.

  • Every thought says how it was made.From "my idea, in my words" to "an agent wrote it", so an agent's draft never passes for a person's conclusion.

  • Disagreement is kept, never scored."3 agree, 1 disagrees, and here is who." A score would bury the minority view.

  • Any surface the person chooses.Their own agent as a reading companion ("what's new since I last looked?"), or the browser as a repository of everything read.

  • Plain files in git.An agent sees only its prompt, the repo and its tools' output. Google's Open Knowledge Format (OKF) describes nearly the same shape, so I aligned our field names with it and skipped the full spec.

03

Where the research stands

What reading together surfaced

Sources that seem opposed often speak from different positions, a creator's or a user's; silent AI actions are the main failure in AI interfaces; and admitting AI use costs trust, but being found out costs more.

The direction

Jakob Nielsen's Intent by Discovery: interfaces are moving from commands to stated intent, with the user as supervisor.

Research questions, draft 3 (draft 2 drew feedback, so each is now one question, with every term defined and no change assumed):

  • AgencyHow do learners exercise and experience control over their learning when an AI shapes the learning environment?

  • SignalsWhich learner behaviours can AI learning tools treat as evidence of learning, and how does that evidence shape their interventions?

  • InterfacesHow do peer and guided learning in a course differ when learners use one shared, institution-provided interface versus interfaces they each modify or create?

Next

A literature review, a way to measure each question, and a first study: an audit of who makes each learning decision in shipped products, from Sana to Khanmigo.

04

Reflection

LEAD, ARCHIE and LIVE are cousins, and together they are my reflection: all three stem from harness engineering and collaborative work between people and agents.

We curate and raise our agents.

LLMs are trained, and agents are trained too, by the harness we build around them, so that tomorrow we can swap the brain, the LLM, or run more complex agentic systems as a side help for research and discovery.