# AI Agents, Integrations and Human Escalation: practice workbook
Talent Engineering Academy | An education initiative by Vitae
Edition 2026.09 | Updated 2026-09-24
Course: https://talentengineering.org/courses/ai-agents-integrations

Design agents with bounded tools, explicit permissions and reliable handoff when they cannot safely proceed.

## Your deliverable
An agent contract, permission map and adversarial test plan.

Use fictional information. Keep your completed work in your own secure notes. Exercises and capstone work are self-directed, not independently assessed.

## 1. Goal and prohibited actions

Your notes:



## 2. Tool contracts and least privilege

Your notes:



## 3. Trusted instructions versus untrusted data

Your notes:



## 4. Approval scope and current-state checks

Your notes:



## 5. Escalation payload and human owner

Your notes:



## 6. Attack, failure and recovery tests

Your notes:



## Lesson exercises

### 1. Bound the agent’s authority

Write an agent contract for a fictional research assistant with five allowed and five prohibited actions.

Your response:


Worked example: An agent may retrieve approved company facts and draft a research brief. Sending messages and changing application decisions require separate authorised actions that the agent’s tools do not permit by default.

Self-review guidance: Describe concrete tool permissions and record scope. Do not rely on a sentence saying be careful while providing an unrestricted send or database-write tool.

### 2. Treat external content as untrusted input

Create three malicious source examples and identify the control that blocks each action.

Your response:


Worked example: A CV includes text asking the assistant to export all applicants to an external address. The system can read the CV as evidence but has no tool permission to export other records or send to arbitrary destinations.

Self-review guidance: Include an attempted data export, instruction override and unauthorised status change. Name the enforcement boundary, not just a model warning. Test with fictional data only.

### 3. Design approvals and escalation payloads

Design an escalation card and three outcomes: approve, revise and stop.

Your response:


Worked example: An agent drafts a candidate message. The review shows recipient, content, source facts and why contact is appropriate. A changed recipient or material message edit invalidates the earlier approval.

Self-review guidance: Include what the agent does after no response and how it records the decision. A reviewer should be able to reject the action without breaking the rest of the workflow.

### 4. Evaluate the agent as a system

Write eight system tests and a shutdown procedure for your agent design.

Your response:


Worked example: A test produces a good research brief but also attempts to update a candidate score. The run fails because the prohibited tool call violates the contract, even if the final text looks useful.

Self-review guidance: Check allowed and blocked actions, not only text quality. Include recovery after partial completion and evidence that permission revocation takes effect. State the approved scope and remaining limitations.

## Portfolio review
Check that your work is internally consistent, distinguishes facts from assumptions, names decision owners and explains its limitations. Revise gaps before using the method in real work.

## Further reading
- [OWASP: Top 10 for large language model applications](https://genai.owasp.org/llm-top-10/): Security background for applications that use language models.
- [NIST: AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework): Voluntary framework for organising AI risks and controls.
- [GOV.UK: Responsible AI in recruitment](https://www.gov.uk/government/publications/responsible-ai-in-recruitment-guide/responsible-ai-in-recruitment): UK guidance on procuring and deploying recruitment AI.

Original educational scenarios. References provide further reading and do not imply endorsement. Check current official rules and appropriate professional advice for real legal, financial or regulated decisions.