AI Workbench · Level 8 of 15
A safe agent
Level 8 of 15 in the AI Workbench: a free AI exercise at Advanced level, which takes about 30 minutes. Build a tool-using flow that completes a refund within limits, needs approval before sending, and does not obey an instruction planted in a file.
Free · no download · about 30 minutes. Your work is saved on this device. Updated 18 Oct 2026.
The brief
- Client
- Brightleaf Home (fictional, hosted by OFORO)
- Site
- OFORO agent desk
- Work order
- OFR-WO-108
- Data set
- Refund ticket
- Task code
- AIW-08
- Difficulty
- Advanced
- Type of work
- Agents
Build a tool-using flow that completes a refund within limits, needs approval before sending, and does not obey an instruction planted in a file.
Skills you practise
- Tool calling
- Limits and budgets
- Human approval
- Prompt injection defence
What the test checks
The acceptance test marks the skills this level practises.
- Tool calling
- Limits and budgets
- Human approval
- Prompt injection defence
In this lab
- Train models in your browser
- Vision inspection, prompts, RAG and safe agents
- Check the AI's work before you trust it
Levels in the AI Workbench
Previous: Answers with sourcesNext: Fairness check
- Data health check
- Train and evaluate without leakage
- Choose the metric and the threshold
- Overfitting: tune the tree
- Vision inspection on the conveyor
- Prompt to spec
- Answers with sources
- A safe agent
- Fairness check
- Privacy first
- Check the AI's work: finance
- Check the AI's work: HR
- Check the AI's work: marketing
- Check the AI's work: healthcare
- Check the AI's work: project management
Part of the AI Workbench
Prepare data, train a model and check the AI's work against a brief.
Feeds these internships: AI and Machine Learning, Data Science and Analytics, Generative AI and Prompt Engineering, AI Agents and Automation, Computer Vision, AI for Finance and Accounting, AI for HR and Recruitment, AI for Digital Marketing and Content, AI in Healthcare and Pharma, AI for Project Management.