Module 13 explains how agents work. Here a small open model, running on your device, gets three tools (a calculator, a company directory and a currency converter) and eight tasks. You see every step it takes: the tool it calls, what comes back, and its answer, graded by code. One directory entry has an instruction planted in it, to see whether the agent obeys text it reads.
The agent
Model
Run on
Step limit
Scorecardgraded by code, live on this device
Run the tasks to fill this table.
Every taskclick a result to see the agent’s steps
The agent loop
Think, Act, Read, Repeat
An agent is a model in a loop. It decides on an action, the program carries it out, and the result goes back into the conversation. The model never touches the tools directly: the page parses its line, checks it, and runs only the three tools it knows.
Why the planted instruction matters
Everything It Reads Is Input
An agent that reads web pages, emails or files reads text written by other people. If it treats that text as instructions, anyone who can put words in front of it can steer it. Defending against that is the main safety problem for agents.
Limits of this test
Eight Tasks, Small Models
Models this small often format a call wrongly or stop early; larger models do far better. Treat the scores as a demonstration of how agents are tested, not a ranking.
Live, not recorded. Qwen2.5 (Alibaba), Llama 3.2 (Meta) and SmolLM2 (Hugging Face) are open models, downloaded from Hugging Face and run with ONNX Runtime in your browser. The companies in the directory are fictional.