Web Navigation Agent on MiniWoB Plus Plus

Can a small language model navigate web pages as an agent? This project builds a web navigation agent on Microsoft AutoGen AgentChat: a planner writes a numbered plan from the task and a compact list of page elements, a navigator turns each plan step into one tool call (click, type_text, select_option, press), a validator checks every action before a headless Playwright browser executes it, and after a failed attempt a reflector writes a lesson for one retry (Reflexion). The pages are MiniWoB++ practice tasks served from 127.0.0.1 by an own environment that blocks every other URL. The agent is compared with a random policy, hand- written scripted heuristics, a single-shot LLM and two ablations on 16 task families x 4 seeds. qwen2.5:3b was served by Ollama on a second laptop (RTX 2050 4 GB, 32 GB RAM) over Wi-Fi; the code ran on a laptop with RTX 2050 4 GB (Ryzen 7, 8 GB RAM).

NOTE: Research demo: the agent acts only on local MiniWoB++ practice pages served from 127.0.0.1; it never opens internet pages, submits real forms or downloads files.

Category: Department of Computer Science and Engineering / Agent Development

Agent Development Project AG10 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.