Agentic AI

Описание: February 10th Show-and-Tell Recap: Agentic AI

Kevin Rank, Liza Long, and Joel Gladd presented Webinar 5 in the Idaho AI Catalyst Initiative series on agentic AI. The webinar defines “agentic AI”, provides a brief intro for how it works differently from pre-2025 genAI interactions, then demonstrates how we’re using agents in our own workflows (especially for meeting the April ADA deadline!).

To define “agent,” we used Anthropic's definition because it's precise. An agent is a system where the LLM directs its own processes and tool usage. A chatbot waits for your prompt, responds, and stops. An agent takes a high-level goal and figures out the steps on its own. It acts, evaluates, revises, and loops until the task is done. Kevin walked through two agentic browsers, Perplexity Comet and Chrome with Gemini. Both can read pages, fill out forms, summarize across tabs, and complete tasks autonomously. Kevin found that agents could even complete a McGraw Hill SimNet Excel simulation. Liza demonstrated using the Claude browser extension to remediate Pressbooks content for ADA compliance ahead of the April 24th Title II deadline. Her workflow fixes heading styles, adds alt text, cleans up Creative Commons attributions, and corrects non-descriptive links, while emphasizing to “always be the human in the loop.” I showed how agentic CLIs like Codex and Claude Code use "skills" (think custom GPTs but at the terminal level) to pull Canvas and Pressbooks content via API, evaluate it for accessibility, and push fixes back.

We also talked about the risks. Agents continuously take screenshots, which creates FERPA concerns. Prompt injection is an ongoing problem. And existing AI policies weren't written for tools that can complete a student's Canvas assignments without the student ever reading them. The slides provide a few resources for thinking about “AI Governance,” including a recent whitepaper and Anna Mills’s substack on agentic browsers.