Evaluate the agent as a system
Course overview · 4 min reading + 12 min practice, estimated
Principles and method
Test tool selection, argument validity, permission enforcement, state handling and final outputs. A correct answer with an unauthorised side effect is a failure. Include long-running tasks, contradictory sources, unavailable tools and malicious content. Trace actions with safe logs that support investigation without storing unnecessary sensitive payloads. Evaluate recovery when the agent stops midway and verify that repeated runs do not duplicate external effects. Monitor scope drift and supplier changes after release. An agent should have a named operator and an immediate way to revoke action permissions.
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.
Put it into practice
Write eight system tests and a shutdown procedure for your agent design.
Use fictional information and keep your work in your own notes.
Compare your approach: 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.
Sources and further reading
Original Academy teaching and fictional examples. These references provide context, not endorsement. Edition 2026.09; updated 2026-09-24.
- OWASP: Top 10 for large language model applications
Security background for applications that use language models.
- NIST: AI Risk Management Framework
Voluntary framework for organising AI risks and controls.
- GOV.UK: Responsible AI in recruitment
UK guidance on procuring and deploying recruitment AI.