Question 4LLMs, RAG and agents

When would you use an agent that takes actions, and how do you keep it safe?

Agent work is growing fast in FDE postings. Senior candidates are expected to know where agents fail.

Asked for in 80% of live FDE postings

“Applied AI Knowledge: Large language models (LLMs), prompt engineering, frameworks like RAG and function calling, and building evals to validate agentic AI solutions.”

Forward Deployed Architect · DevRev

They scoreStarts from cost of mistakesLimits what it can touchHumans approve risky steps

The answerFour steps

1
Use agents for multi-step tasks with clear success checks; avoid them where one wrong action is costly
Say it like this
“I'd use an agent where the task has several steps and a result you can check, like refund requests: read the order, check the policy, issue the refund.”
Why it scores
Picks agents by checkable outcome and cost of a wrong action, and says where not to use one.
2
Give the agent narrow tools and permissions, not broad access
Say it like this
“It gets three tools: look up an order, read the policy, refund up to the order total. No database access, no admin token.”
Why it scores
Narrow tools with hard limits in code. The prompt is not the permission system.
3
Require human approval for risky actions
Say it like this
“Refunds over $200, or a second refund on the same order, wait for a person to approve.”
Why it scores
Draws the approval line with a number and knows what it costs in human time.
4
Log every step so the customer can audit it
Say it like this
“Every tool call is logged with its input, result and the model's stated reason, so their finance team can audit any refund in a minute.”
Why it scores
An audit trail the customer can use. Security teams ask for this first.

StrongAll together

About a minute, out loud. The marked parts are the four steps; hover one to find it above.

I'd use an agent where the task has several steps and a result you can check, like refund requests: read the order, check the policy, issue the refund.¹ I wouldn't use one to change prices across a catalog, where one wrong action hits thousands of customers. It gets three tools: look up an order, read the policy, refund up to the order total. No database access, no admin token.² Refunds over $200, or a second refund on the same order, wait for a person to approve.³ Say that's 8% of requests, so the team still saves most of its time. Every tool call is logged with its input, result and the model's stated reason, so their finance team can audit any refund in a minute.⁴

WeakThe answer that fails

Sounds fine. Common. Scores low.

“Agents are great for automating workflows. I'd give it access to the systems it needs, write a strong system prompt saying what it must never do, and monitor it closely.”

Interviewer's note

Safety lives in the prompt. Broad access plus “don't do bad things” fails on the first odd input.