Over the past few months I've been building an agentic teacher's assistant that handles a significant amount of my day-to-day tasks.
This tool connects to my school's learning management system and acts as a chatbot that is loaded with 83 atomic tools and skills like: planning and posting lessons, assignments, assessments, rubrics. It can create, read, and update any content on Canvas.
It can quickly grade student work and provide feedback that is more helpful and comprehensive than what I can consistently produce by myself.
I have a pedagogical council of AI advisors that help me plan lessons, projects, and assessments. The council is made up of three voices patterned after Seymour Papert, John Dewey, and Maria Montessori. Papert pushes me toward constructionism and learning by making. Dewey keeps me honest about experience and inquiry. Montessori watches for student agency and the prepared environment. I give them a draft lesson, and the council reads it, argues with it, and hands back a revised version.
It knows all about Franklin's core philosophy and organizes learning around nine Transdisciplinary Competencies - skills that cut across every subject and define what it means to be a Franklin graduate. The assistant has a skill called the Serendipity Generator that can look across the curriculum of all my colleagues and identify transdisciplinary project opportunities that align to student interests.
All of this is done in one chat experience. There is no copying or pasting. Once I approve my assistant's draft, I have it pushed to Canvas. It cannot publish to Canvas. That is a design choice. I want to ensure that I do a final check on Canvas before I push the publish button and make any content viewable by students. All student names and identifiers are replaced with anonymous IDs before anything hits the large language model. This tool is FERPA compliant.
Because I built this as an AI Agent with small tools and skills plus access to our learning management system, it can combine those tools and skills to express emergent capabilities that it was not originally designed for.
When one of our science teachers unexpectedly had to leave mid-year, I was able to generate a comprehensive transition report for administration and the substitute teacher about where her classes were, what had been taught, what hadn't been taught, how each individual student was doing.
When the grade level deans needed a daily report that shows what students got D's or F's on any assignments that day, I was able to spin it up with a few prompts and a Slack bot that now automates that task.
The Department Leads used to spend hours each week checking every Canvas page and assessment to check for compliance with a list of Franklin guidelines. I was able to ask my tool to create and automate that report. So now they receive a Google Doc at the end of each week that flags any issues on Canvas.
When we want to get a full picture of how a particular student is doing on their nine Transdisciplinary Competencies, I can search out all assignments and assessments that the student has ever done to curate a portfolio of work that demonstrates or fails to demonstrate those skills.
I love building with AI and for AI Agents. In January, I started with a few small tools that could accomplish specific tasks. I linked the tools together and learned about Agent Native Architecture.
I realized that I was growing software rather than building it.
I am planting seeds and wondering: what will this become?